CN117808127B - Image processing method, federated learning method and device under data heterogeneity conditions - Google Patents

Abstract

The present invention discloses an image processing method, a federated learning method and a device under data heterogeneity conditions, and relates to the field of image processing technology. Edge computing devices are clustered according to the similarity of data distribution. The edge computing devices in the cluster have similar data distribution, which can allow the model to better capture the characteristics of the data and effectively solve the problem of data heterogeneity. The edge computing devices in the cluster aggregate model parameters according to the tree-shaped aggregation network in the cluster. The edge computing devices in the lower layer only send model parameters to the corresponding edge computing devices in the upper layer, and do not send model parameters to other edge computing devices, which can greatly reduce communication overhead. The edge computing device and the edge cloud server perform two-layer model parameter aggregation in the federated learning process to obtain an accurate and reliable image processing model. Finally, the edge computing device uses the accurate and reliable image processing model to perform image processing, which can improve the accuracy and reliability of image processing.

Description

Image processing method, federated learning method and device under data heterogeneity conditions

Technical Field

The present invention relates to the field of image processing technology, and in particular to an image processing method, a federated learning method, a device, a system, a device and a medium under data heterogeneity conditions.

Background technique

Federated learning is a distributed machine learning method. The participants in federated learning do not exchange local data, but local model parameters. Local model parameters refer to the model parameters obtained by edge computing devices using local data to train the model. The edge cloud server aggregates the local model parameters of each edge computing device to obtain the global model parameters, and sends the global model to each edge computing device. The global model is a model with global model parameters as model parameters. However, traditional federated learning schemes often do not consider the differences in data distribution between edge computing devices, which will lead to poor adaptability of the model to the data distribution of some edge computing devices. For federated learning environments with large data heterogeneity, this will cause the performance of the model to be impaired. In addition, traditional federated learning schemes usually use a centralized approach to aggregate model parameters, that is, all edge computing devices send their own model parameters to the edge cloud server for aggregation, which will bring about a large communication overhead. In the case where the trained global model is an image processing model, the accuracy and reliability of image processing using the trained image processing model are not high.

In view of this, how to solve the problem of data heterogeneity in federated learning, reduce communication overhead, and improve the accuracy and reliability of image processing has become a technical problem that technical personnel in this field need to solve urgently.

Summary of the invention

The purpose of the present invention is to provide an image processing method, a federated learning method, an apparatus, a system, a device and a medium under data heterogeneity conditions, which can solve the problem of data heterogeneity in federated learning and reduce communication overhead.

In order to solve the above technical problems, the present invention provides an image processing method under data heterogeneity conditions, comprising:

According to the data distribution similarity of the edge computing devices, the edge computing devices are divided into a plurality of data homogeneity clusters;

Selecting an edge computing device from the data homogeneity cluster as a cluster head of the data homogeneity cluster;

Receive the in-cluster model parameter aggregation result of the data homogeneity cluster uploaded by the cluster head; the in-cluster model parameter aggregation result is the first-layer model parameter aggregation result obtained by each edge computing device in the data homogeneity cluster aggregating local model parameters according to the in-cluster tree aggregation network, each edge computing device in the data homogeneity cluster constructs the in-cluster tree aggregation network according to the optimal communication strategy, the cluster head is the root node of the in-cluster tree aggregation network, and the child nodes in the in-cluster tree aggregation network send model parameters to the corresponding parent nodes, and perform model parameter aggregation with the local model parameters of the parent nodes;

Aggregate the intra-cluster model parameter aggregation results of each of the data homogeneity clusters to obtain a second-layer model parameter aggregation result, and when the image processing model using the second-layer model parameter aggregation result as the model parameter converges, send the image processing model to each of the edge computing devices, so that the edge computing device uses the image processing model to process the image.

In some embodiments, selecting an edge computing device from the data homogeneity cluster as a cluster head of the data homogeneity cluster includes:

An edge computing device in the data homogeneity cluster that is closest to other edge computing devices in the cluster or has the highest communication rate with other edge computing devices in the cluster is selected as the cluster head of the data homogeneity cluster.

In some embodiments, the cluster head selects a preset number of edge computing devices from other edge computing devices in the cluster as the first-layer child nodes of the tree-shaped aggregation network within the cluster, and the first-layer child nodes are edge computing devices with the highest data transmission rate when communicating with the cluster head; the i-th layer child nodes select a preset number of edge computing devices from the remaining edge computing devices in the cluster as the i+1-th layer child nodes of the tree-shaped aggregation network within the cluster; the i+1-th layer child nodes are edge computing devices with the highest data transmission rate when communicating with the i-th layer child nodes; i takes values from 1 until the tree-shaped aggregation network within the cluster is constructed.

In some embodiments, aggregating the intra-cluster model parameter aggregation results of each of the data homogeneity clusters to obtain the second-layer model parameter aggregation results includes:

Determine the weight coefficient of each data homogeneity cluster in the current training round;

According to the weight coefficients of the data homogeneity clusters, the in-cluster model parameters of the data homogeneity clusters are weightedly aggregated to obtain the aggregation result of the second-layer model parameters of the current training round.

In some embodiments, determining the weight coefficient of each data homogeneity cluster in the current training round includes:

The local data test accuracy of each edge computing device of the data homogeneity cluster is counted to obtain the local data test accuracy of the data homogeneity cluster; the local data test accuracy of the edge computing device is the data test accuracy of the edge computing device using the local data to test the global model obtained in the previous training round;

The weight coefficient of the data homology cluster is determined according to the local data test accuracy of the data homology cluster; wherein the weight coefficient of the data homology cluster is negatively correlated with the local data test accuracy of the data homology cluster.

In some embodiments, determining the weight coefficient of the data homogeneity cluster according to the local data test accuracy of the data homogeneity cluster includes:

according to The weight coefficient of the data homology cluster is obtained; K _c represents the weight coefficient of the cth data homology cluster, and μ represents the local data test accuracy of the cth data homology cluster.

In some embodiments, a training round includes three stages: local model training, first-layer model parameter aggregation, and second-layer model parameter aggregation; after each edge computing device in the data homogeneity cluster completes a first preset number of model parameter updates, a first-layer model parameter aggregation is performed; after each data homogeneity cluster completes a second preset number of first-layer model parameter aggregations, a second-layer model parameter aggregation is performed.

In some embodiments, it also includes:

Whenever the preset training rounds of model training are completed, the edge computing devices are clustered again according to the data distribution similarity of the edge computing devices, and the preset training rounds of model training are performed again until the global model converges.

In some embodiments, according to the data distribution similarity of the edge computing devices, dividing the edge computing devices into a plurality of data homogeneity clusters includes:

According to the data distribution similarity of each edge computing device, a weighted undirected graph is constructed; the value of the connecting edge between two edge computing devices in the weighted undirected graph is the value of the data distribution similarity of the two edge computing devices;

According to the weighted undirected graph, the edge computing device is divided into a plurality of data homogeneity clusters.

In some embodiments, constructing a weighted undirected graph according to the data distribution similarity of each edge computing device includes:

Searching for public data from the public network, and building a test data set based on the public data;

Sending the test data set to each edge computing device so that each edge computing device uses a local model to infer the test data set;

Receiving the inference results uploaded by each edge computing device, and calculating the similarity of each inference result;

The weighted undirected graph is constructed according to the similarity of the inference results.

In some embodiments, constructing the weighted undirected graph according to the similarity of the inference results includes:

Compare the similarity of the inference results of each edge computing device with the preset threshold;

If the similarity of the inference results of the two edge computing devices is greater than a preset threshold, a connection relationship between the two edge computing devices is established, and the value of the similarity of the inference results of the two edge computing devices is used as the value of the connection edge between the two edge computing devices.

In some embodiments, according to the weighted undirected graph, dividing the edge computing device into a plurality of data homogeneity clusters includes:

Initialize the label of each edge computing device in the weighted undirected graph;

Iteratively updating the label of each edge computing device; wherein iteratively updating the label of each edge computing device includes: setting the label of the edge computing device connected to the connection edge whose value is greater than the set threshold to the same label; counting the number of times the label of the neighbor edge computing device of the target edge computing device appears, and selecting the label with the largest number of appearances among the neighbor edge computing devices as the label of the target edge computing device; the target edge computing device is an edge computing device to which the values of each connection edge connected are not greater than the set threshold;

Determine whether the iterative update stop condition is met;

If the iterative update stop condition is met, the edge computing devices with the same label are divided into the same data homogeneity cluster.

In some embodiments, determining whether the iterative update stop condition is satisfied includes:

After each iteration, the change between the label of each edge computing device after the current iteration and the label of each edge computing device after the previous iteration is calculated.

Comparing the change amount with a preset value;

If the change amount is less than the preset value, the iterative update stop condition is met.

In order to solve the above technical problems, the present invention also provides a federated learning method under data heterogeneity conditions, including:

According to the data distribution similarity of the edge computing devices, the edge computing devices are divided into a plurality of data homogeneity clusters;

Selecting an edge computing device from the data homogeneity cluster as a cluster head of the data homogeneity cluster;

Receive the in-cluster model parameter aggregation result of the data homogeneity cluster uploaded by the cluster head; the in-cluster model parameter aggregation result is the first-layer model parameter aggregation result obtained by each edge computing device in the data homogeneity cluster aggregating local model parameters according to the in-cluster tree aggregation network, each edge computing device in the data homogeneity cluster constructs the in-cluster tree aggregation network according to the optimal communication strategy, the cluster head is the root node of the in-cluster tree aggregation network, and the child nodes in the in-cluster tree aggregation network send model parameters to the corresponding parent nodes, and perform model parameter aggregation with the local model parameters of the parent nodes;

Aggregate the intra-cluster model parameter aggregation results of each of the data homogeneity clusters to obtain a second-layer model parameter aggregation result, and when a global model using the second-layer model parameter aggregation result as a model parameter converges, send the global model to each of the edge computing devices.

In order to solve the above technical problems, the present invention further provides an image processing device under data heterogeneity conditions, comprising:

A partitioning module, configured to partition the edge computing devices into a plurality of data homogeneity clusters according to the similarity of data distribution of the edge computing devices;

A selection module, configured to select an edge computing device from the data homogeneity cluster as a cluster head of the data homogeneity cluster;

A receiving module is used to receive the intra-cluster model parameter aggregation result of the data homogeneity cluster uploaded by the cluster head; the intra-cluster model parameter aggregation result is the first-layer model parameter aggregation result obtained by each edge computing device in the data homogeneity cluster aggregating local model parameters according to the intra-cluster tree aggregation network, each edge computing device in the data homogeneity cluster constructs the intra-cluster tree aggregation network according to the optimal communication strategy, the cluster head is the root node of the intra-cluster tree aggregation network, and the child nodes in the intra-cluster tree aggregation network send model parameters to the corresponding parent nodes, and perform model parameter aggregation with the local model parameters of the parent nodes;

An aggregation module is used to aggregate the intra-cluster model parameter aggregation results of each of the data homogeneity clusters to obtain a second-layer model parameter aggregation result, and when the image processing model using the second-layer model parameter aggregation result as the model parameter converges, the image processing model is sent to each of the edge computing devices, so that the edge computing device uses the image processing model to process the image.

In order to solve the above technical problems, the present invention further provides a federated learning device under data heterogeneity conditions, comprising:

A division unit, configured to divide the edge computing devices into a plurality of data homogeneity clusters according to the data distribution similarity of the edge computing devices;

A selection unit, configured to select an edge computing device from the data homogeneity cluster as a cluster head of the data homogeneity cluster;

A receiving unit is used to receive the intra-cluster model parameter aggregation result of the data homogeneity cluster uploaded by the cluster head; the intra-cluster model parameter aggregation result is a first-layer model parameter aggregation result obtained by each edge computing device in the data homogeneity cluster aggregating local model parameters according to the intra-cluster tree aggregation network, each edge computing device in the data homogeneity cluster constructs the intra-cluster tree aggregation network according to the optimal communication strategy, the cluster head is the root node of the intra-cluster tree aggregation network, and the child nodes in the intra-cluster tree aggregation network send model parameters to the corresponding parent nodes, and perform model parameter aggregation with the local model parameters of the parent nodes;

An aggregation unit is used to aggregate the intra-cluster model parameter aggregation results of each of the data homogeneity clusters to obtain a second-layer model parameter aggregation result, and when a global model using the second-layer model parameter aggregation result as a model parameter converges, the global model is sent to each of the edge computing devices.

In order to solve the above technical problems, the present invention further provides a device, comprising:

Memory for storing computer programs;

A processor is used to implement the steps of the image processing method under the data heterogeneity condition as described above or the steps of the federated learning method under the data heterogeneity condition as described above when executing the computer program.

In order to solve the above technical problems, the present invention further provides an image processing system under data heterogeneity conditions, comprising:

Edge computing devices are used to train image processing models using local data, obtain local model parameters, and process images using image processing modules sent by edge cloud servers;

The edge cloud server is used to divide the edge computing devices into several data homogeneity clusters according to the data distribution similarity of the edge computing devices; select an edge computing device from the data homogeneity cluster as the cluster head of the data homogeneity cluster; receive the cluster model parameter aggregation result of the data homogeneity cluster uploaded by the cluster head; the cluster model parameter aggregation result is the first layer model parameter aggregation result obtained by each edge computing device in the data homogeneity cluster aggregating local model parameters according to the cluster tree aggregation network, each edge computing device in the data homogeneity cluster constructs the cluster tree aggregation network according to the optimal communication strategy, the cluster head is the root node of the cluster tree aggregation network, and the child nodes in the cluster tree aggregation network send model parameters to the corresponding parent nodes, and perform model parameter aggregation with the local model parameters of the parent nodes; aggregate the cluster model parameter aggregation results of each data homogeneity cluster to obtain a second layer model parameter aggregation result, and when the image processing model using the second layer model parameter aggregation result as the model parameter converges, send the image processing model to each edge computing device, so that the edge computing device uses the image processing model to process the image.

In order to solve the above technical problems, the present invention further provides a federated learning system under data heterogeneity conditions, including:

Edge computing devices are used to train models using local data and obtain local model parameters;

The edge cloud server is used to divide the edge computing devices into several data homogeneity clusters according to the data distribution similarity of the edge computing devices; select an edge computing device from the data homogeneity cluster as the cluster head of the data homogeneity cluster; receive the cluster model parameter aggregation result of the data homogeneity cluster uploaded by the cluster head; the cluster model parameter aggregation result is the first layer model parameter aggregation result obtained by each edge computing device in the data homogeneity cluster aggregating local model parameters according to the cluster tree aggregation network, each edge computing device in the data homogeneity cluster constructs the cluster tree aggregation network according to the optimal communication strategy, the cluster head is the root node of the cluster tree aggregation network, the child nodes in the cluster tree aggregation network send model parameters to the corresponding parent nodes, and perform model parameter aggregation with the local model parameters of the parent nodes; aggregate the cluster model parameter aggregation results of each data homogeneity cluster to obtain the second layer model parameter aggregation result, and when the global model with the second layer model parameter aggregation result as the model parameter converges, the global model is sent to each edge computing device.

In order to solve the above technical problems, the present invention also provides a medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the image processing method under the above-mentioned data heterogeneous conditions or the above-mentioned federated learning method under the above-mentioned data heterogeneous conditions are implemented.

The image processing method under data heterogeneity provided by the present invention clusters edge computing devices according to the similarity of data distribution. The edge computing devices in the cluster have similar data distribution. Since similar data distribution allows the model to better capture the characteristics of the data, the aggregation of model parameters in the cluster can improve the accuracy of model training for each cluster and effectively solve the problem of data heterogeneity. In addition, the edge computing devices in the cluster aggregate model parameters according to the tree-shaped aggregation network in the cluster, and the child nodes only send model parameters to the corresponding parent nodes, that is, the edge computing devices in the lower layer only send model parameters to the corresponding edge computing devices in the upper layer, and do not send model parameters to other edge computing devices, which can greatly reduce communication overhead. At the same time, after the edge computing devices are clustered, if the edge computing devices in a cluster fail or go offline, other clusters can still perform local model training and model parameter aggregation normally, thereby improving the overall fault tolerance and scalability of the federated training system. The edge computing devices and the edge cloud server perform two-layer model parameter aggregation in the federated learning process to obtain an accurate and reliable image processing model. Finally, the edge computing devices use the accurate and reliable image processing model for image processing, which can improve the accuracy and reliability of image processing.

The federated learning method under data heterogeneity provided by the present invention clusters edge computing devices according to the similarity of data distribution. The edge computing devices in the cluster have similar data distribution. Since similar data distribution allows the model to better capture the characteristics of the data, the aggregation of model parameters in the cluster can improve the accuracy of model training in each cluster and effectively solve the problem of data heterogeneity. In addition, the edge computing devices in the cluster aggregate model parameters according to the tree-shaped aggregation network in the cluster. The child nodes only send model parameters to the corresponding parent nodes, that is, the edge computing devices in the lower layer only send model parameters to the corresponding edge computing devices in the upper layer, and do not send model parameters to other edge computing devices, which can greatly reduce communication overhead. At the same time, after the edge computing devices are clustered, if the edge computing devices in a cluster fail or go offline, other clusters can still perform local model training and model parameter aggregation normally, thereby improving the overall fault tolerance and scalability of the federated training system.

The image processing device, federated learning device, equipment, system and medium under data heterogeneity conditions provided by the present invention all have the above-mentioned technical effects.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the prior art and the drawings required for use in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

FIG1 is a schematic diagram of a flow chart of a federated learning method under heterogeneous data conditions provided by an embodiment of the present invention;

FIG2 is a schematic diagram of a weighted undirected graph provided by an embodiment of the present invention;

FIG3 is a schematic diagram of a clustering result provided by an embodiment of the invention;

FIG4 is a schematic diagram of an intra-cluster tree aggregation network provided by an embodiment of the present invention;

FIG5 is a schematic diagram of a federated learning device under data heterogeneity conditions provided by an embodiment of the present invention;

FIG6 is a schematic flow chart of an image processing method under heterogeneous data conditions provided by an embodiment of the present invention;

FIG. 7 is a schematic diagram of an image processing device under data heterogeneity conditions provided by an embodiment of the present invention.

Detailed ways

The core of the present invention is to provide an image processing method, a federated learning method, an apparatus, a system, a device and a medium under data heterogeneity conditions, which can solve the problem of data heterogeneity in federated learning, reduce communication overhead, and improve the accuracy and reliability of image processing.

In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

Please refer to FIG. 1 , which is a flow chart of a federated learning method under heterogeneous data conditions provided by an embodiment of the present invention. Referring to FIG. 1 , the method includes:

S101: Divide the edge computing devices into a plurality of data homogeneity clusters according to data distribution similarity of the edge computing devices;

The data distribution of edge computing devices within the same data homogeneity cluster is similar.

In some embodiments, dividing the edge computing devices into a plurality of data homogeneity clusters according to data distribution similarity of the edge computing devices includes:

According to the data distribution similarity of each edge computing device, a weighted undirected graph is constructed; the value of the connecting edge between two edge computing devices in the weighted undirected graph is the value of the data distribution similarity of the two edge computing devices;

According to the weighted undirected graph, the edge computing device is divided into a plurality of data homogeneity clusters.

This embodiment constructs a weighted undirected graph based on the data distribution similarity of edge computing devices, and then divides the edge computing devices into a number of data homogeneity clusters based on the weighted undirected graph. The value of the connecting edge between two edge computing devices in the weighted undirected graph is the value of the data distribution similarity of the two edge computing devices.

In some embodiments, constructing a weighted undirected graph according to the data distribution similarity of each edge computing device includes:

Searching for public data from the public network, and building a test data set based on the public data;

Sending the test data set to each edge computing device so that each edge computing device uses a local model to infer the test data set;

Receiving the inference results uploaded by each edge computing device, and calculating the similarity of each inference result;

The weighted undirected graph is constructed according to the similarity of the inference results.

Each edge computing device performs model training based on local data to obtain a local model. The edge cloud server searches for public data from the public network and builds a test data set based on the public data. The edge cloud server sends the test data set to each edge computing device. The edge computing device stores the test data set and uses the local model trained based on the local data to infer the test data set to obtain the inference result. Each edge computing device uploads the inference result to the edge cloud server. The edge cloud server calculates the similarity of each inference result and builds a weighted undirected graph of each edge computing device based on the similarity of the inference result.

The edge cloud server can use a vector similarity calculation method to calculate the similarity of each inference result. Exemplarily, the edge cloud server uses the Jaccard similarity coefficient calculation method to calculate the similarity of each inference result. The Jaccard similarity coefficient calculation can be used to calculate the similarity between sets, and can also be used to calculate the similarity of binary vectors. For two binary vectors A and B, the calculation formula for the Jaccard similarity coefficient calculation is: similarity = |A ∩ B| / |A ∪ B|; wherein A ∩ B represents the intersection of vectors A and B, and A ∪ B represents the union of vectors A and B.

Assume that the inference result of edge computing device C is a binary vector [1,0,0,0,…1,1,1,0], and the inference result of edge computing device D is a binary vector [0,1,1,0,…1,1,1,0]. Use the Jaccard similarity coefficient calculation method to calculate the similarity between the binary vector [1,0,0,0,…1,1,1,0] and the binary vector [0,1,1,0,…1,1,1,0].

This embodiment characterizes the similarity of data distribution between edge computing devices by the similarity of the inference results of the edge computing devices for the same test data set. The more similar the inference results of the edge computing devices using local models to infer the test data set, the more similar the local models trained by the edge computing devices using local data are. The more similar the local models trained by the edge computing devices using local data are, the more similar the data distribution of the edge computing devices is. Using the data homogeneity clustering method provided in this embodiment, the edge computing devices only need to upload the inference results instead of uploading local data, which can ensure the privacy of local data.

In some embodiments, constructing the weighted undirected graph according to the similarity of the inference results includes:

Compare the similarity of the inference results of each edge computing device with the preset threshold;

If the similarity of the inference results of the two edge computing devices is greater than a preset threshold, a connection relationship between the two edge computing devices is established, and the value of the similarity of the inference results of the two edge computing devices is used as the value of the connection edge between the two edge computing devices.

If the similarity of the inference results of two edge computing devices is greater than a preset threshold, a connection relationship between the two edge computing devices is established, and the value of the connection edge between the two edge computing devices is equal to the value of the similarity of the inference results between the two edge computing devices. If the similarity of the inference results of the two edge computing devices is not greater than the preset threshold, a connection relationship between the two edge computing devices is not established. Exemplarily, the preset threshold is set to 0.7, and the weighted undirected graph shown in Figure 2 is constructed accordingly. In Figure 2, devices 1 to 6 represent six edge computing devices, and the value on the connection edge between two devices is the value of the similarity of the prediction results of the two devices.

In some embodiments, according to the weighted undirected graph, dividing the edge computing device into a plurality of data homogeneity clusters includes:

Initialize the label of each edge computing device in the weighted undirected graph;

Iteratively updating the label of each edge computing device; wherein iteratively updating the label of each edge computing device includes: setting the label of the edge computing device connected to the connection edge whose value is greater than the set threshold to the same label; counting the number of times the label of the neighbor edge computing device of the target edge computing device appears, and selecting the label with the largest number of appearances among the neighbor edge computing devices as the label of the target edge computing device; the target edge computing device is an edge computing device to which the values of each connection edge connected are not greater than the set threshold;

Determine whether the iterative update stop condition is met;

If the iterative update stop condition is met, the edge computing devices with the same label are divided into the same data homogeneity cluster.

The labels of each edge computing device in the weighted undirected graph are initialized to different labels. The edge cloud server iteratively updates the labels of each edge computing device. The process of one iterative update includes: comparing the size of each connection edge with the set threshold, and setting the labels of the edge computing devices connected to the connection edges whose values are greater than the set threshold to the same label. For edge computing devices whose values of each connection edge are not greater than the set threshold, the number of times the labels of the neighboring edge computing devices of the edge computing device appear is further counted, and the label with the most appearances in the neighboring edge computing devices is selected as the label of the edge computing device. Edge computing devices with a connection relationship in a weighted undirected graph are neighbors of each other. When the iterative update stop condition is met, the edge computing devices with the same label are divided into the same data homogeneity cluster, and the same data homogeneity cluster is a collection of edge computing devices with the same label.

Exemplarily, referring to the weighted undirected graph shown in FIG2, assuming that the threshold is set to 0.9, the value of the connection edge between device 1 and device 2 is 0.94>0.9, so the labels of device 1 and device 2 are set to the same label, for example, set to label A. The value of the connection edge between device 3 and device 4 is 0.91>0.9, so the labels of device 3 and device 4 are set to the same label, for example, set to label B. The value of the connection edge between device 5 and device 4 and the value of the connection edge between device 5 and device 3 are not greater than 0.9, so the labels of the neighbor edge computing devices (devices 3 and 4) of device 5 are counted. Since the labels of device 3 and device 4 are both label B, and device 5 has no other neighbor edge computing devices, the label that appears most frequently in the neighbor edge computing devices of device 5 is label B (appears twice), so label B is used as the label of device 5. The value of the connection edge between device 6 and device 4 and the value of the connection edge between device 6 and device 3 are not greater than 0.9, so the labels of the neighbor edge computing devices (devices 3 and 4) of device 6 are counted. Since the labels of device 3 and device 4 are both label B, and device 6 has no other neighbor edge computing devices, the label that appears most frequently among the neighbor edge computing devices of device 6 is label B (appears twice), so label B is used as the label of device 6.

After the iterative update stop condition is met, the labels of device 1 and device 2 are both label A, and the labels of device 3 to device 6 are all label B. Therefore, device 1 and device 2 are divided into the same data homogeneity cluster, and device 3 to device 6 are divided into the same data homogeneity cluster, and the clustering result shown in Figure 3 is obtained.

In some embodiments, the determining whether the iterative update stopping condition is satisfied includes:

After each iteration, the change between the label of each edge computing device after the current iteration and the label of each edge computing device after the previous iteration is calculated.

Comparing the change amount with a preset value;

If the change amount is less than the preset value, the iterative update stop condition is met.

After each iteration, the edge cloud server calculates the change between the labels of each edge computing device after this iteration and the labels of each edge computing device after the previous iteration. If the change is less than the preset value, that is, the label is basically stable and no longer changes significantly, the algorithm is considered to have converged and the iteration is stopped. If the change is not less than the preset value, that is, the label is unstable and still changes significantly, the iteration continues.

After clustering is completed, the edge cloud server sends the device number of each edge computing device in the divided data homogeneity cluster to each edge computing device in the data homogeneity cluster, so that the edge computing device can communicate with other edge cloud servers in the same cluster accordingly.

S102: Selecting an edge computing device from the data homogeneity cluster as a cluster head of the data homogeneity cluster;

The cluster head is responsible for uploading the aggregation results of the cluster model parameters to the edge cloud server.

In some embodiments, selecting an edge computing device from the data homogeneity cluster as a cluster head of the data homogeneity cluster includes:

An edge computing device that is closest to other edge computing devices in the data homogeneity cluster or has the highest communication rate with other edge computing devices is selected as the cluster head of the data homogeneity cluster.

In some embodiments, selecting an edge computing device from the data homogeneity cluster as a cluster head of the data homogeneity cluster includes:

An edge computing device in the data homogeneity cluster that is closest to other edge computing devices in the cluster or has the highest communication rate with other edge computing devices in the cluster is selected as the cluster head of the data homogeneity cluster.

In this embodiment, the cluster head is selected based on the principle of maximum communication rate or shortest communication distance. Specifically, the edge computing device that is closest to other edge computing devices in the cluster or has the highest communication rate with other edge computing devices is selected as the cluster head. This can reduce the communication distance and delay and improve communication efficiency. After selecting the cluster head, the edge cloud server sends the device number of the cluster head to the edge computing devices in the cluster.

S103: Receive the intra-cluster model parameter aggregation result of the data homogeneity cluster uploaded by the cluster head; the intra-cluster model parameter aggregation result is the first-layer model parameter aggregation result obtained by each edge computing device in the data homogeneity cluster aggregating local model parameters according to the intra-cluster tree aggregation network, each edge computing device in the data homogeneity cluster constructs the intra-cluster tree aggregation network according to the optimal communication strategy, the cluster head is the root node of the intra-cluster tree aggregation network, and the child nodes in the intra-cluster tree aggregation network send model parameters to the parent node, and perform model parameter aggregation with the local model parameters of the parent node.

As shown in FIG3 , the intra-cluster tree aggregation network includes several layers, and each edge computing device is a node in the intra-cluster tree aggregation network. The cluster head is located at the first layer and is the root node of the intra-cluster tree aggregation network. The child nodes of the cluster head (the layer 1 cluster head shown in FIG3 ) are located at the second layer; the child nodes of the child nodes of the cluster head (the layer 2 cluster head shown in FIG3 ) are located at the third layer; and so on. The layer 1 cluster head is the parent node of the layer 2 cluster head, and the layer 1 cluster head is also the child node of the cluster head. Similarly, the layer 2 cluster head is the parent node of the layer 3 cluster head, and the layer 2 cluster head is also the child node of the layer 1 cluster head. And so on. The layer 1 cluster head only sends model parameters to the cluster head, and the layer 2 cluster head only sends model parameters to the layer 1 cluster head, and so on.

The aggregation method of model parameters within the cluster is as follows: the edge computing device of the Nth layer (the last layer), i.e., the leaf node, uploads its local model parameters obtained by training based on local data to the parent node of the edge computing device located at the N-1th layer. The edge computing device of the N-1th layer aggregates the local model parameters uploaded by its child nodes with its own local model parameters, and uploads the parameter aggregation results to the parent node of the edge computing device located at the N-2th layer. And so on. Finally, the cluster head aggregates the parameter aggregation results uploaded by its child nodes with its own local model parameters to obtain the aggregation results of model parameters within the cluster.

The method for obtaining the parameter aggregation result by aggregating the local model parameters of the non-leaf node with the model parameters uploaded by the child nodes of the non-leaf node can be: the non-leaf node averages the model parameters uploaded by its child nodes with its own local model parameters. For example, the cluster head averages the parameter aggregation result uploaded by its child nodes with its own local model parameters to obtain the model parameter aggregation result within the cluster.

The intra-cluster tree aggregation network is constructed by each edge computing device in the data homogeneity cluster according to the optimal communication strategy.

In some embodiments, the cluster head selects a preset number of edge computing devices from other edge computing devices in the cluster as the first-layer child nodes of the tree-shaped aggregation network within the cluster, and the first-layer child nodes are edge computing devices with the highest data transmission rate when communicating with the cluster head; the i-th layer child nodes select a preset number of edge computing devices from the remaining edge computing devices in the cluster as the i+1-th layer child nodes of the tree-shaped aggregation network within the cluster; the i+1-th layer child nodes are edge computing devices with the highest data transmission rate when communicating with the i-th layer child nodes; i takes values from 1 until the tree-shaped aggregation network within the cluster is constructed.

The edge cloud server selects one from the edge computing devices in the data homogeneity cluster as the cluster head of the data homogeneity cluster. The cluster head of the data homogeneity cluster selects a preset number of edge computing devices with the highest data transmission rate when communicating with the cluster head from the remaining edge computing devices in the cluster as the first-layer child nodes, i.e., the layer 1 cluster head. The layer 1 cluster head selects a preset number of edge computing devices with the highest data transmission rate when communicating with the layer 1 cluster head from the remaining edge computing devices in the cluster as the second-layer child nodes, i.e., the layer 2 cluster head. This is done by analogy until all edge computing devices in the cluster obtain their own numbers in the tree-shaped aggregation network in the cluster, i.e., until each edge computing device in the cluster knows which edge computing device to upload the model parameters to. The cluster head only receives the model parameters of the layer 1 cluster head, but not the model parameters of other edge computing devices in the cluster. The layer 1 cluster head only receives the model parameters of the layer 2 cluster head, but not the model parameters of other edge computing devices in the cluster. By analogy, the layer N-1 cluster head only receives the model parameters of the layer N cluster head, but not the model parameters of other edge computing devices in the cluster.

In the process of building a tree-shaped aggregation network within a cluster, the cluster head communicates with all edge computing devices within the cluster through wireless links, and the data communicated can be model parameters. The cluster head collects the model parameters uploaded by all edge computing devices within the cluster and calculates the sending rate of each edge computing device. The sending rate can be calculated as follows: sending rate = total communication data volume/(total sending time + total receiving time).

The cluster head selects a preset number of edge computing devices with the highest transmission rate as the layer 1 cluster head according to the transmission rate of each edge computing device in the cluster. After receiving the appointment instruction of the cluster head, the layer 1 cluster head communicates with other edge computing devices in the cluster except the cluster head and the layer 1 cluster head, calculates the transmission rate of other edge computing devices in the cluster, and selects the preset number of edge computing devices with the highest transmission rate as the layer 2 cluster head; after receiving the appointment instruction of the layer 1 cluster head, the layer 2 cluster head communicates with other edge computing devices in the cluster except the cluster head, the layer 1 cluster head and the layer 2 cluster head, calculates the transmission rate of other edge computing devices in the cluster, and selects the preset number of edge computing devices with the highest transmission rate as the layer 3 cluster head; and so on.

Exemplarily, the cluster head selects the two edge computing devices with the highest transmission rates as the layer 1 cluster head according to the transmission rates of each edge computing device in the cluster. After receiving the appointment instruction of the cluster head, the layer 1 cluster head communicates with other edge computing devices in the cluster except the cluster head and the layer 1 cluster head, calculates the transmission rates of other edge computing devices in the cluster, and selects the two edge computing devices with the highest transmission rates as the layer 2 cluster head; after receiving the appointment instruction of the layer 1 cluster head, the layer 2 cluster head communicates with other edge computing devices in the cluster except the cluster head, the layer 1 cluster head and the layer 2 cluster head, calculates the transmission rates of other edge computing devices in the cluster, and selects the two edge computing devices with the highest transmission rates as the layer 3 cluster head; and so on.

S104: Aggregate the intra-cluster model parameter aggregation results of each of the data homogeneity clusters to obtain a second-layer model parameter aggregation result, and when a global model using the second-layer model parameter aggregation result as a model parameter converges, send the global model to each of the edge computing devices.

The edge cloud server aggregates the cluster model parameter aggregation results of each data homogeneity cluster to obtain the global model parameters, that is, the second-layer model parameter aggregation results. When the global model with the second-layer model parameter aggregation results as the model parameters converges, the converged global model is sent to each edge computing device. The edge computing device uses the global model to perform corresponding detection, analysis, etc. Exemplarily, if the global model obtained by federated learning is an image processing model, then the edge computing device uses the global model for image processing. If the global model obtained by federated learning is a network attack detection model, then the edge computing device uses the global model for network attack detection.

In some embodiments, aggregating the intra-cluster model parameter aggregation results of each of the data homogeneity clusters to obtain the second-layer model parameter aggregation results includes:

Determine the weight coefficient of each data homogeneity cluster in the current training round;

According to the weight coefficients of the data homogeneity clusters, the in-cluster model parameters of the data homogeneity clusters are weightedly aggregated to obtain the aggregation result of the second-layer model parameters of the current training round.

This embodiment sets a corresponding weight coefficient for each data homogeneity cluster, and performs weighted aggregation on the cluster model parameter aggregation results uploaded by the cluster head of each data homogeneity cluster according to the weight coefficient of each data homogeneity cluster, so that the model can better reflect the real global data distribution and improve the generalization ability of the model.

In some embodiments, determining the weight coefficient of each data homogeneity cluster in the current training round includes:

The local data test accuracy of each edge computing device of the data homogeneity cluster is counted to obtain the local data test accuracy of the data homogeneity cluster; the local data test accuracy of the edge computing device is the data test accuracy of the edge computing device using local data to test the global model obtained in the previous training round;

The weight coefficient of the data homology cluster is determined according to the local data test accuracy of the data homology cluster; wherein the weight coefficient of the data homology cluster is negatively correlated with the local data test accuracy of the data homology cluster.

After the edge cloud server broadcasts the global model obtained in the previous training round to all edge computing devices, each edge computing device uses local data for testing. The testing method can be: the edge computing device randomly selects 1/10 of the local data as test data to test the global model. After the test is completed, the edge computing device sends the test results of the device to the edge cloud server. The edge cloud server traverses each data homogeneity cluster and counts the local data test accuracy of each data homogeneity cluster. The statistical method is to find the average value of the local data test accuracy of each edge computing device in the cluster. If the local data test accuracy of a data homogeneity cluster is low, it means that the data training of the data homogeneity cluster is insufficient or the global model is not suitable for the data of the data homogeneity cluster. At this time, the weight coefficient of the data homogeneity cluster is increased.

This embodiment determines the weight coefficient of the data homogeneity cluster according to the local data test accuracy of the data homogeneity cluster. The local data test accuracy of the data homogeneity cluster is different, and the weight coefficient of the data homogeneity cluster is different. By assigning different weight coefficients to different data homogeneity clusters, it is possible to avoid edge computing devices with large data volume or good quality being ignored. At the same time, the weight coefficient of the data homogeneity cluster is determined according to the local data test accuracy of the data homogeneity cluster. The weight coefficient of the data homogeneity cluster is dynamically adjusted, which can avoid the model from over-adapting to a specific edge computing device and improve the generalization ability of the model.

Among them, in some embodiments, determining the weight coefficient of the data homogeneity cluster according to the local data test accuracy of the data homogeneity cluster includes:

according to The weight coefficient of the data homogeneity cluster is obtained; K _c represents the weight coefficient of the cth data homogeneity cluster, and μ represents the local data test accuracy of the cth data homogeneity cluster. /> The base is e and the exponent is -μ.

In some embodiments, a training round includes three stages: local model training, first-layer model parameter aggregation, and second-layer model parameter aggregation; after each edge computing device in the data homogeneity cluster completes a first preset number of model parameter updates, a first-layer model parameter aggregation is performed; after each data homogeneity cluster completes a second preset number of first-layer model parameter aggregations, a second-layer model parameter aggregation is performed.

Assume that all edge computing devices are divided into C data homogeneous clusters, consisting of the set Indicates that the kth data homogeneity cluster S _k contains/> edge computing devices. In the federated learning system, edge computing device i trains a local model based on its own local data set _Di. The local empirical loss function of the data distribution at edge computing device i is:/> ; Among them, /> are local model parameters, /> is the data sample involved in iterative training,/> is a sample loss function used to quantify data samples/> The prediction error on .

The main goal of federated learning is to optimize the global model parameters to minimize the global loss function associated with all edge computing devices while ensuring the highest aggregation efficiency. The global loss function is:

.

The model training process in federated learning is divided into three stages: local model training, first-layer model parameter aggregation, i.e., cluster-wide model parameter aggregation, and second-layer model parameter aggregation, i.e., global aggregation. These three stages are combined into a training round.

Local model training: Each edge computing device updates the local model. In some embodiments, the edge computing devices in the cluster use the SGD (Stochastic gradient descent) algorithm to update the local model. During the tth round of training, the lth iteration update process is expressed as:

; Among them, /> is the updated local model parameter in the tth round and the lth iteration,/> is the learning rate of the lth iteration of the tth round.

In-cluster model parameter aggregation: The edge computing devices in the cluster perform After the iteration update, the cluster model parameter aggregation is performed. The aggregation method can be: the layer k cluster head receives the model parameters uploaded by its child nodes and averages them with its own local model parameters, and the cluster head receives the model parameters uploaded by its child nodes and averages them with its own local model parameters to obtain the cluster model parameter aggregation result. Among them, the edge computing devices in the cluster are updated every / > After training, the model is stored. When receiving the aggregation instruction sent by the upper cluster head, the stored model is sent to the upper cluster head connected to it.

Global aggregation: After all data homogeneous clusters perform intra-cluster model parameter aggregation τ times, the edge cloud server performs global aggregation in a synchronous manner. The edge cloud server receives the model parameters uploaded by C cluster heads and updates the global model parameters by parameter averaging:

; /> is the model parameter of the cth cluster in the tth training round,/> is the global model parameter updated in the tth training round.

After the global model parameters are updated, the edge cloud server broadcasts the global model to all edge computing devices, and each edge computing device uses local data for testing. The testing method can be: the edge computing device randomly selects 1/10 of the local data as test data to test the global model. After the test is completed, the edge computing device sends the test results of the device to the edge cloud server. The edge cloud server traverses each data homogeneity cluster and counts the local data test accuracy of each data homogeneity cluster. The statistical method is to find the average value of the local data test accuracy of each edge computing device in the cluster.

The global aggregation for the next training round is expressed as:

; /> The model parameters of the cth cluster in the t+1th training round,/> is the global model parameter updated in the t+1th training round.

, K _c represents the weight coefficient of the c-th data homogeneity cluster, and μ represents the local data test accuracy of the c-th data homogeneity cluster.

In some embodiments, it also includes:

Whenever the preset training rounds of model training are completed, the edge computing devices are clustered again according to the data distribution similarity of the edge computing devices, and the preset training rounds of model training are performed again until the global model converges.

After completing the model training of h training rounds, re-cluster the edge computing devices into data homogeneity clusters, and perform model training again after re-clustering the data homogeneity clusters (including local model training, model parameter aggregation within the cluster, and global aggregation). Repeat the above steps of dividing the data homogeneity clusters and performing model training after dividing the data homogeneity clusters until the global model converges.

In summary, the federated learning method under data heterogeneity provided by the present invention clusters edge computing devices according to the similarity of data distribution. The edge computing devices in the cluster have similar data distribution. Since similar data distribution allows the model to better capture the characteristics of the data, the aggregation of model parameters in the cluster can improve the accuracy of model training in each cluster and effectively solve the problem of data heterogeneity. In addition, the edge computing devices in the cluster aggregate model parameters according to the tree-shaped aggregation network in the cluster. The child nodes only send model parameters to the corresponding parent nodes, that is, the edge computing devices in the lower layer only send model parameters to the corresponding edge computing devices in the upper layer, and do not send model parameters to other edge computing devices. This can greatly reduce communication overhead. At the same time, after the edge computing devices are clustered, if the edge computing device in a cluster fails or goes offline, other clusters can still perform local model training and model parameter aggregation normally, thereby improving the overall fault tolerance and scalability of the federated training system.

The present invention also provides a federated learning device under data heterogeneity conditions, and the device described below can correspond to the method described above. Please refer to Figure 5, which is a schematic diagram of a federated learning device under data heterogeneity conditions provided by an embodiment of the present invention. As shown in Figure 5, the device includes:

A division unit 11 is used to divide the edge computing devices into a plurality of data homogeneity clusters according to the data distribution similarity of the edge computing devices;

A selection unit 12, configured to select an edge computing device from the data homogeneity cluster as a cluster head of the data homogeneity cluster;

The receiving unit 13 is used to receive the intra-cluster model parameter aggregation result of the data homogeneity cluster uploaded by the cluster head; the intra-cluster model parameter aggregation result is the first-layer model parameter aggregation result obtained by each edge computing device in the data homogeneity cluster aggregating local model parameters according to the intra-cluster tree aggregation network, each edge computing device in the data homogeneity cluster constructs the intra-cluster tree aggregation network according to the optimal communication strategy, the cluster head is the root node of the intra-cluster tree aggregation network, and the child nodes in the intra-cluster tree aggregation network send model parameters to the corresponding parent nodes, and perform model parameter aggregation with the local model parameters of the parent nodes;

The aggregation unit 14 is used to aggregate the intra-cluster model parameter aggregation results of each of the data homogeneity clusters to obtain the second-layer model parameter aggregation results, and when the global model using the second-layer model parameter aggregation results as the model parameters converges, the global model is sent to each of the edge computing devices.

In some embodiments, the selection unit 12 is specifically configured to:

An edge computing device in the data homogeneity cluster that is closest to other edge computing devices in the cluster or has the highest communication rate with other edge computing devices in the cluster is selected as the cluster head of the data homogeneity cluster.

In some embodiments, the cluster head selects a preset number of edge computing devices from other edge computing devices in the cluster as the first-layer child nodes of the tree-shaped aggregation network within the cluster, and the first-layer child nodes are edge computing devices with the highest data transmission rate when communicating with the cluster head; the i-th layer child nodes select a preset number of edge computing devices from the remaining edge computing devices in the cluster as the i+1-th layer child nodes of the tree-shaped aggregation network within the cluster; the i+1-th layer child nodes are edge computing devices with the highest data transmission rate when communicating with the i-th layer child nodes; i takes values from 1 until the tree-shaped aggregation network within the cluster is constructed.

In some embodiments, the polymerization unit 14 includes:

A determination subunit, used to determine the weight coefficient of each data homogeneity cluster in the current training round;

The aggregation subunit is used to perform weighted aggregation on the intra-cluster model parameters of each data homogeneity cluster according to the weight coefficient of each data homogeneity cluster to obtain the second-layer model parameter aggregation result of the current training round.

In some embodiments, the determining subunit is specifically configured to:

The local data test accuracy of each edge computing device of the data homogeneity cluster is counted to obtain the local data test accuracy of the data homogeneity cluster; the local data test accuracy of the edge computing device is the data test accuracy of the edge computing device using local data to test the global model obtained in the previous training round;

The weight coefficient of the data homology cluster is determined according to the local data test accuracy of the data homology cluster; wherein the weight coefficient of the data homology cluster is negatively correlated with the local data test accuracy of the data homology cluster.

In some embodiments, the determining subunit is specifically configured to:

according to The weight coefficient of the data homology cluster is obtained; K _c represents the weight coefficient of the cth data homology cluster, and μ represents the local data test accuracy of the cth data homology cluster.

In some embodiments, a training round includes three stages: local model training, first-layer model parameter aggregation, and second-layer model parameter aggregation; after each edge computing device in the data homogeneity cluster completes a first preset number of model parameter updates, a first-layer model parameter aggregation is performed; after each data homogeneity cluster completes a second preset number of first-layer model parameter aggregations, a second-layer model parameter aggregation is performed.

In some embodiments, it also includes:

The repetition unit is used to re-cluster the edge computing devices according to the data distribution similarity of the edge computing devices after completing the model training of the preset training rounds, and perform the model training of the preset training rounds again until the global model converges.

In some embodiments, the dividing unit 11 includes:

A construction subunit is used to construct a weighted undirected graph according to the data distribution similarity of each edge computing device; the value of the connecting edge between two edge computing devices in the weighted undirected graph is the value of the data distribution similarity of the two edge computing devices;

The partitioning subunit is used to partition the edge computing device into a plurality of data homogeneous clusters according to the weighted undirected graph.

In some embodiments, the building block includes:

A search subunit, used to search for public data from the public network and construct a test data set based on the public data;

A sending subunit, configured to send the test data set to each edge computing device, so that each edge computing device uses a local model to infer the test data set;

A computing subunit, configured to receive the inference results uploaded by each edge computing device and calculate the similarity of each inference result;

The weighted undirected graph construction subunit is used to construct the weighted undirected graph according to the similarity of the reasoning results.

In some embodiments, the weighted undirected graph construction subunit includes:

A comparison subunit, used to compare the similarity of the inference results of each edge computing device with a preset threshold;

A subunit is established to establish a connection relationship between the two edge computing devices if the similarity of the inference results of the two edge computing devices is greater than a preset threshold, and the value of the similarity of the inference results of the two edge computing devices is used as the value of the connection edge between the two edge computing devices.

In some embodiments, dividing the subunits includes:

An initialization subunit, used to initialize the label of each edge computing device in the weighted undirected graph;

An iterative update subunit, configured to iteratively update the labels of each edge computing device; wherein iteratively updating the labels of each edge computing device includes: setting the labels of edge computing devices connected to connection edges whose values are greater than a set threshold to the same label; counting the number of times the labels of neighboring edge computing devices of the target edge computing device appear, and selecting the label with the largest number of appearances among the neighboring edge computing devices as the label of the target edge computing device; the target edge computing device is an edge computing device to which the values of each connection edge connected are not greater than the set threshold;

A judgment subunit, used to judge whether the iterative update stop condition is met;

The clustering subunit is used to divide the edge computing devices with the same label into the same data homogeneity cluster if the iterative update stop condition is met.

In some embodiments, the determination subunit is specifically used for:

After each iteration, the change between the label of each edge computing device after the current iteration and the label of each edge computing device after the previous iteration is calculated.

Comparing the change amount with a preset value;

If the change amount is less than the preset value, the iterative update stop condition is met.

The present invention also provides a federated learning system under data heterogeneity conditions, including: an edge computing device and an edge cloud server.

Edge computing devices are used to train models using local data and obtain local model parameters;

The edge cloud server is used to divide the edge computing devices into several data homogeneity clusters according to the data distribution similarity of the edge computing devices; select an edge computing device from the data homogeneity cluster as the cluster head of the data homogeneity cluster; receive the cluster model parameter aggregation result of the data homogeneity cluster uploaded by the cluster head; the cluster model parameter aggregation result is the first layer model parameter aggregation result obtained by each edge computing device in the data homogeneity cluster aggregating local model parameters according to the cluster tree aggregation network, each edge computing device in the data homogeneity cluster constructs the cluster tree aggregation network according to the optimal communication strategy, the cluster head is the root node of the cluster tree aggregation network, the child nodes in the cluster tree aggregation network send model parameters to the corresponding parent nodes, and perform model parameter aggregation with the local model parameters of the parent nodes; aggregate the cluster model parameter aggregation results of each data homogeneity cluster to obtain the second layer model parameter aggregation result, and when the global model with the second layer model parameter aggregation result as the model parameter converges, the global model is sent to each edge computing device.

For an introduction to the system provided by the present invention, please refer to the above method embodiment, and the present invention will not be elaborated here.

Please refer to FIG. 6 , which is a flow chart of an image processing method under data heterogeneity conditions provided by an embodiment of the present invention. Referring to FIG. 6 , the method includes:

S201: Divide the edge computing devices into a plurality of data homogeneity clusters according to data distribution similarity of the edge computing devices;

S202: Selecting an edge computing device from the data homogeneity cluster as a cluster head of the data homogeneity cluster;

S203: Receive the cluster model parameter aggregation result of the data homogeneity cluster uploaded by the cluster head; the cluster model parameter aggregation result is the first layer model parameter aggregation result obtained by each edge computing device in the data homogeneity cluster aggregating local model parameters according to the cluster tree aggregation network, each edge computing device in the data homogeneity cluster constructs the cluster tree aggregation network according to the optimal communication strategy, the cluster head is the root node of the cluster tree aggregation network, and the child nodes in the cluster tree aggregation network send model parameters to the corresponding parent nodes, and perform model parameter aggregation with the local model parameters of the parent nodes;

S204: Aggregate the intra-cluster model parameter aggregation results of each of the data homogeneity clusters to obtain a second-layer model parameter aggregation result, and when the image processing model using the second-layer model parameter aggregation result as the model parameter converges, send the image processing model to each of the edge computing devices, so that the edge computing device uses the image processing model to process the image.

In some embodiments, selecting an edge computing device from the data homogeneity cluster as a cluster head of the data homogeneity cluster includes:

An edge computing device in the data homogeneity cluster that is closest to other edge computing devices in the cluster or has the highest communication rate with other edge computing devices in the cluster is selected as the cluster head of the data homogeneity cluster.

In some embodiments, the cluster head selects a preset number of edge computing devices from other edge computing devices in the cluster as the first-layer child nodes of the tree-shaped aggregation network within the cluster, and the first-layer child nodes are edge computing devices with the highest data transmission rate when communicating with the cluster head; the i-th layer child nodes select a preset number of edge computing devices from the remaining edge computing devices in the cluster as the i+1-th layer child nodes of the tree-shaped aggregation network within the cluster; the i+1-th layer child nodes are edge computing devices with the highest data transmission rate when communicating with the i-th layer child nodes; i takes values from 1 until the tree-shaped aggregation network within the cluster is constructed.

In some embodiments, aggregating the intra-cluster model parameter aggregation results of each of the data homogeneity clusters to obtain the second-layer model parameter aggregation results includes:

Determine the weight coefficient of each data homogeneity cluster in the current training round;

According to the weight coefficients of the data homogeneity clusters, the in-cluster model parameters of the data homogeneity clusters are weightedly aggregated to obtain the aggregation result of the second-layer model parameters of the current training round.

In some embodiments, determining the weight coefficient of each data homogeneity cluster in the current training round includes:

The local data test accuracy of each edge computing device of the data homogeneity cluster is counted to obtain the local data test accuracy of the data homogeneity cluster; the local data test accuracy of the edge computing device is the data test accuracy of the edge computing device using the local data to test the global model obtained in the previous training round;

The weight coefficient of the data homology cluster is determined according to the local data test accuracy of the data homology cluster; wherein the weight coefficient of the data homology cluster is negatively correlated with the local data test accuracy of the data homology cluster.

In some embodiments, determining the weight coefficient of the data homogeneity cluster according to the local data test accuracy of the data homogeneity cluster includes:

according to The weight coefficient of the data homology cluster is obtained; K _c represents the weight coefficient of the cth data homology cluster, and μ represents the local data test accuracy of the cth data homology cluster.

In some embodiments, a training round includes three stages: local model training, first-layer model parameter aggregation, and second-layer model parameter aggregation; after each edge computing device in the data homogeneity cluster completes a first preset number of model parameter updates, a first-layer model parameter aggregation is performed; after each data homogeneity cluster completes a second preset number of first-layer model parameter aggregations, a second-layer model parameter aggregation is performed.

In some embodiments, it also includes:

Whenever the preset training rounds of model training are completed, the edge computing devices are clustered again according to the data distribution similarity of the edge computing devices, and the preset training rounds of model training are performed again until the global model converges.

In some embodiments, dividing the edge computing devices into a plurality of data homogeneity clusters according to data distribution similarity of the edge computing devices includes:

According to the data distribution similarity of each edge computing device, a weighted undirected graph is constructed; the value of the connecting edge between two edge computing devices in the weighted undirected graph is the value of the data distribution similarity of the two edge computing devices;

According to the weighted undirected graph, the edge computing device is divided into a plurality of data homogeneity clusters.

In some embodiments, constructing a weighted undirected graph according to the data distribution similarity of each edge computing device includes:

Searching for public data from the public network, and building a test data set based on the public data;

Sending the test data set to each edge computing device so that each edge computing device uses a local model to infer the test data set;

Receiving the inference results uploaded by each edge computing device, and calculating the similarity of each inference result;

The weighted undirected graph is constructed according to the similarity of the inference results.

In some embodiments, constructing the weighted undirected graph according to the similarity of the inference results includes:

Compare the similarity of the inference results of each edge computing device with the preset threshold;

If the similarity of the inference results of the two edge computing devices is greater than a preset threshold, a connection relationship between the two edge computing devices is established, and the value of the similarity of the inference results of the two edge computing devices is used as the value of the connection edge between the two edge computing devices.

In some embodiments, according to the weighted undirected graph, dividing the edge computing device into a plurality of data homogeneity clusters includes:

Initialize the label of each edge computing device in the weighted undirected graph;

Iteratively updating the label of each edge computing device; wherein iteratively updating the label of each edge computing device includes: setting the label of the edge computing device connected to the connection edge whose value is greater than the set threshold to the same label; counting the number of times the label of the neighbor edge computing device of the target edge computing device appears, and selecting the label with the largest number of appearances among the neighbor edge computing devices as the label of the target edge computing device; the target edge computing device is an edge computing device to which the values of each connection edge connected are not greater than the set threshold;

Determine whether the iterative update stop condition is met;

If the iterative update stop condition is met, the edge computing devices with the same label are divided into the same data homogeneity cluster.

In some embodiments, determining whether the iterative update stop condition is satisfied includes:

After each iteration, the change between the label of each edge computing device after the current iteration and the label of each edge computing device after the previous iteration is calculated.

Comparing the change amount with a preset value;

If the change amount is less than the preset value, the iterative update stop condition is met.

In summary, the image processing method under data heterogeneity provided by the present invention clusters edge computing devices according to the similarity of data distribution. The edge computing devices in the cluster have similar data distribution. Since similar data distribution allows the model to better capture the characteristics of the data, the aggregation of model parameters in the cluster can improve the accuracy of model training for each cluster and effectively solve the problem of data heterogeneity. In addition, the edge computing devices in the cluster perform model parameter aggregation according to the tree-shaped aggregation network in the cluster, and the child nodes only send model parameters to the corresponding parent nodes, that is, the edge computing devices in the lower layer only send model parameters to the corresponding edge computing devices in the upper layer, and do not send model parameters to other edge computing devices, which can greatly reduce communication overhead. At the same time, after the edge computing devices are clustered, if the edge computing devices in a cluster fail or go offline, other clusters can still perform local model training and model parameter aggregation normally, thereby improving the overall fault tolerance and scalability of the federated training system. The edge computing devices and the edge cloud server perform two-layer model parameter aggregation in the federated learning process to obtain an accurate and reliable image processing model. Finally, the edge computing devices use the accurate and reliable image processing model for image processing, which can improve the accuracy and reliability of image processing.

The present invention also provides an image processing device under data heterogeneity conditions. The device described below can correspond to the method described above. Please refer to Figure 7, which is a schematic diagram of an image processing device under data heterogeneity conditions provided by an embodiment of the present invention. As shown in Figure 7, the device includes:

A division module 21, configured to divide the edge computing devices into a plurality of data homogeneity clusters according to the data distribution similarity of the edge computing devices;

A selection module 22, configured to select an edge computing device from the data homogeneity cluster as a cluster head of the data homogeneity cluster;

The receiving module 23 is used to receive the intra-cluster model parameter aggregation result of the data homogeneity cluster uploaded by the cluster head; the intra-cluster model parameter aggregation result is the first-layer model parameter aggregation result obtained by each edge computing device in the data homogeneity cluster aggregating local model parameters according to the intra-cluster tree aggregation network, each edge computing device in the data homogeneity cluster constructs the intra-cluster tree aggregation network according to the optimal communication strategy, the cluster head is the root node of the intra-cluster tree aggregation network, and the child nodes in the intra-cluster tree aggregation network send model parameters to the corresponding parent nodes, and perform model parameter aggregation with the local model parameters of the parent nodes;

The aggregation module 24 is used to aggregate the intra-cluster model parameter aggregation results of each of the data homogeneity clusters to obtain the second-layer model parameter aggregation results, and when the image processing model using the second-layer model parameter aggregation results as model parameters converges, the image processing model is sent to each of the edge computing devices, so that the edge computing device uses the image processing model to process the image.

In some embodiments, the selection module 22 is specifically used to:

An edge computing device in the data homogeneity cluster that is closest to other edge computing devices in the cluster or has the highest communication rate with other edge computing devices in the cluster is selected as the cluster head of the data homogeneity cluster.

In some embodiments, the cluster head selects a preset number of edge computing devices from other edge computing devices in the cluster as the first-layer child nodes of the tree-shaped aggregation network within the cluster, and the first-layer child nodes are edge computing devices with the highest data transmission rate when communicating with the cluster head; the i-th layer child nodes select a preset number of edge computing devices from the remaining edge computing devices in the cluster as the i+1-th layer child nodes of the tree-shaped aggregation network within the cluster; the i+1-th layer child nodes are edge computing devices with the highest data transmission rate when communicating with the i-th layer child nodes; i takes values from 1 until the tree-shaped aggregation network within the cluster is constructed.

In some embodiments, the aggregation module 24 includes:

A determination submodule is used to determine the weight coefficients of the data homogeneity clusters in the current training round;

The aggregation submodule is used to perform weighted aggregation on the intra-cluster model parameters of each data homogeneity cluster according to the weight coefficient of each data homogeneity cluster to obtain the second-layer model parameter aggregation result of the current training round.

In some embodiments, the determination submodule is specifically configured to:

The local data test accuracy of each edge computing device of the data homogeneity cluster is counted to obtain the local data test accuracy of the data homogeneity cluster; the local data test accuracy of the edge computing device is the data test accuracy of the edge computing device using local data to test the global model obtained in the previous training round;

The weight coefficient of the data homology cluster is determined according to the local data test accuracy of the data homology cluster; wherein the weight coefficient of the data homology cluster is negatively correlated with the local data test accuracy of the data homology cluster.

In some embodiments, the determination submodule is specifically configured to:

according to The weight coefficient of the data homology cluster is obtained; K _c represents the weight coefficient of the cth data homology cluster, and μ represents the local data test accuracy of the cth data homology cluster.

In some embodiments, a training round includes three stages: local model training, first-layer model parameter aggregation, and second-layer model parameter aggregation; after each edge computing device in the data homogeneity cluster completes a first preset number of model parameter updates, a first-layer model parameter aggregation is performed; after each data homogeneity cluster completes a second preset number of first-layer model parameter aggregations, a second-layer model parameter aggregation is performed.

In some embodiments, it also includes:

The repetition module is used to re-cluster the edge computing devices according to the data distribution similarity of the edge computing devices after completing the model training of the preset training rounds, and perform the model training of the preset training rounds again until the global model converges.

In some embodiments, the partitioning module 21 includes:

A construction submodule is used to construct a weighted undirected graph according to the data distribution similarity of each edge computing device; the value of the connecting edge between two edge computing devices in the weighted undirected graph is the value of the data distribution similarity of the two edge computing devices;

The partitioning submodule is used to partition the edge computing device into a plurality of data homogeneous clusters according to the weighted undirected graph.

In some embodiments, the building blocks include:

A search submodule, used to search for public data from the public network and construct a test data set based on the public data;

A sending submodule, used for sending the test data set to each edge computing device, so that each edge computing device uses a local model to infer the test data set;

A calculation submodule, used to receive the inference results uploaded by each edge computing device, and calculate the similarity of each inference result;

The weighted undirected graph construction submodule is used to construct the weighted undirected graph according to the similarity of the reasoning results.

In some embodiments, the weighted undirected graph construction submodule includes:

A comparison submodule is used to compare the similarity of the inference results of each edge computing device with a preset threshold;

A submodule is established to establish a connection relationship between the two edge computing devices if the similarity of the inference results of the two edge computing devices is greater than a preset threshold, and the value of the similarity of the inference results of the two edge computing devices is used as the value of the connection edge between the two edge computing devices.

In some embodiments, the partitioning submodule includes:

An initialization submodule, used to initialize the labels of each edge computing device in the weighted undirected graph;

An iterative update submodule, for iteratively updating the labels of each edge computing device; wherein iteratively updating the labels of each edge computing device includes: setting the labels of edge computing devices connected to the connection edges whose values are greater than a set threshold to the same label; counting the number of times the labels of neighboring edge computing devices of the target edge computing device appear, and selecting the label with the largest number of appearances among the neighboring edge computing devices as the label of the target edge computing device; the target edge computing device is an edge computing device to which the values of each connected connection edge are not greater than the set threshold;

A judgment submodule is used to judge whether the iterative update stop condition is met;

The clustering submodule is used to divide the edge computing devices with the same label into the same data homogeneity cluster if the iterative update stop condition is met.

In some embodiments, the determination submodule is specifically used for:

After each iteration, the change between the label of each edge computing device after the current iteration and the label of each edge computing device after the previous iteration is calculated.

Comparing the change amount with a preset value;

If the change amount is less than the preset value, the iterative update stop condition is met.

The present invention also provides a device, which includes a memory and a processor.

Memory for storing computer programs;

A processor is used to execute a computer program to implement the steps of any embodiment of the federated learning method or the steps of any embodiment of the image processing method.

For an introduction to the device provided by the present invention, please refer to the above method embodiment, and the present invention will not be elaborated here.

The present invention also provides a medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of any embodiment of the federated learning method or the steps of any embodiment of the image processing method can be implemented.

The medium may include: a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program codes.

For an introduction to the medium provided by the present invention, please refer to the above method embodiment, and the present invention will not be elaborated here.

The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other. For the devices, equipment and media disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method part description.

Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented directly using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of medium known in the art.

The above is a detailed introduction to the federated learning method, image processing method, and device under data heterogeneity conditions provided by the present invention. This article uses specific examples to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present invention, the present invention can also be improved and modified in a number of ways, and these improvements and modifications also fall within the scope of protection of the present invention.

Claims (19)

1. An image processing method under heterogeneous data conditions, characterized by comprising: According to the data distribution similarity of the edge computing devices, the edge computing devices are divided into a plurality of data homogeneity clusters; Selecting an edge computing device in the data homogeneity cluster that is closest to other edge computing devices in the cluster or has the highest communication rate with other edge computing devices in the cluster as a cluster head of the data homogeneity cluster; Receive the in-cluster model parameter aggregation result of the data homogeneity cluster uploaded by the cluster head; the in-cluster model parameter aggregation result is the first-layer model parameter aggregation result obtained by each edge computing device in the data homogeneity cluster aggregating local model parameters according to the in-cluster tree aggregation network, each edge computing device in the data homogeneity cluster constructs the in-cluster tree aggregation network according to the optimal communication strategy, the cluster head is the root node of the in-cluster tree aggregation network, and the child nodes in the in-cluster tree aggregation network send model parameters to the corresponding parent nodes, and perform model parameter aggregation with the local model parameters of the parent nodes; Aggregate the intra-cluster model parameter aggregation results of each of the data homogeneity clusters to obtain a second-layer model parameter aggregation result, and when the image processing model using the second-layer model parameter aggregation result as the model parameter converges, send the image processing model to each of the edge computing devices, so that the edge computing device uses the image processing model to process the image. 2. The image processing method according to claim 1 is characterized in that the cluster head selects a preset number of edge computing devices from other edge computing devices in the cluster as the first-layer child nodes of the tree-shaped aggregation network within the cluster, and the first-layer child nodes are edge computing devices with the largest data transmission rate when communicating with the cluster head; the i-th layer child nodes select a preset number of edge computing devices from the remaining edge computing devices in the cluster as the i+1-th layer child nodes of the tree-shaped aggregation network within the cluster; the i+1-th layer child nodes are edge computing devices with the largest data transmission rate when communicating with the i-th layer child nodes; i takes values from 1 until the tree-shaped aggregation network within the cluster is constructed. 3. The image processing method according to claim 2, characterized in that aggregating the intra-cluster model parameter aggregation results of each of the data homogeneity clusters to obtain the second-layer model parameter aggregation results comprises: Determine the weight coefficient of each data homogeneity cluster in the current training round; According to the weight coefficients of the data homogeneity clusters, the in-cluster model parameters of the data homogeneity clusters are weightedly aggregated to obtain the aggregation result of the second-layer model parameters of the current training round. 4. The image processing method according to claim 3, characterized in that determining the weight coefficient of each data homogeneity cluster in the current training round comprises: The local data test accuracy of each edge computing device of the data homogeneity cluster is counted to obtain the local data test accuracy of the data homogeneity cluster; the local data test accuracy of the edge computing device is the data test accuracy of the edge computing device using the local data to test the global model obtained in the previous training round; The weight coefficient of the data homology cluster is determined according to the local data test accuracy of the data homology cluster; wherein the weight coefficient of the data homology cluster is negatively correlated with the local data test accuracy of the data homology cluster. 5. The image processing method according to claim 4, characterized in that determining the weight coefficient of the data homogeneity cluster according to the local data test accuracy of the data homogeneity cluster comprises: according to The weight coefficient of the data homology cluster is obtained; K _c represents the weight coefficient of the cth data homology cluster, and μ represents the local data test accuracy of the cth data homology cluster. 6. The image processing method according to claim 5 is characterized in that one training round includes three stages: local model training, first-layer model parameter aggregation and second-layer model parameter aggregation; after each edge computing device in the data homogeneity cluster completes the first preset number of model parameter updates, a first-layer model parameter aggregation is performed, and after each data homogeneity cluster completes the second preset number of first-layer model parameter aggregation, a second-layer model parameter aggregation is performed. 7. The image processing method according to claim 6, further comprising: Whenever the preset training rounds of model training are completed, the edge computing devices are clustered again according to the data distribution similarity of the edge computing devices, and the preset training rounds of model training are performed again until the global model converges. 8. The image processing method according to claim 1, characterized in that dividing the edge computing devices into a plurality of data homogeneity clusters according to the data distribution similarity of the edge computing devices comprises: According to the data distribution similarity of each edge computing device, a weighted undirected graph is constructed; the value of the connecting edge between two edge computing devices in the weighted undirected graph is the value of the data distribution similarity of the two edge computing devices; According to the weighted undirected graph, the edge computing device is divided into a plurality of data homogeneous clusters. 9. The image processing method according to claim 8, characterized in that constructing a weighted undirected graph according to the data distribution similarity of each edge computing device comprises: Searching for public data from the public network, and building a test data set based on the public data; Sending the test data set to each edge computing device so that each edge computing device uses a local model to infer the test data set; Receiving the inference results uploaded by each edge computing device, and calculating the similarity of each inference result; The weighted undirected graph is constructed according to the similarity of the inference results. 10. The image processing method according to claim 9, characterized in that constructing the weighted undirected graph according to the similarity of the inference results comprises: Compare the similarity of the inference results of each edge computing device with the preset threshold; If the similarity of the inference results of the two edge computing devices is greater than a preset threshold, a connection relationship between the two edge computing devices is established, and the value of the similarity of the inference results of the two edge computing devices is used as the value of the connection edge between the two edge computing devices. 11. The image processing method according to claim 10, characterized in that, according to the weighted undirected graph, dividing the edge computing device into a plurality of data homogeneity clusters comprises: Initialize the label of each edge computing device in the weighted undirected graph; Iteratively updating the label of each edge computing device; wherein iteratively updating the label of each edge computing device includes: setting the label of the edge computing device connected to the connection edge whose value is greater than the set threshold to the same label; counting the number of times the label of the neighbor edge computing device of the target edge computing device appears, and selecting the label with the largest number of appearances among the neighbor edge computing devices as the label of the target edge computing device; the target edge computing device is an edge computing device to which the values of each connection edge connected are not greater than the set threshold; Determine whether the iterative update stop condition is met; If the iterative update stop condition is met, the edge computing devices with the same label are divided into the same data homogeneity cluster. 12. The image processing method according to claim 11, wherein determining whether the iterative update stop condition is satisfied comprises: After each iteration, the change between the label of each edge computing device after the current iteration and the label of each edge computing device after the previous iteration is calculated. Comparing the change amount with a preset value; If the change amount is less than the preset value, the iterative update stop condition is met. 13. A federated learning method under heterogeneous data conditions, characterized by comprising: According to the data distribution similarity of the edge computing devices, the edge computing devices are divided into a plurality of data homogeneity clusters; Selecting an edge computing device from the data homogeneity cluster as a cluster head of the data homogeneity cluster; Receive the in-cluster model parameter aggregation result of the data homogeneity cluster uploaded by the cluster head; the in-cluster model parameter aggregation result is the first-layer model parameter aggregation result obtained by each edge computing device in the data homogeneity cluster aggregating local model parameters according to the in-cluster tree aggregation network, each edge computing device in the data homogeneity cluster constructs the in-cluster tree aggregation network according to the optimal communication strategy, the cluster head is the root node of the in-cluster tree aggregation network, and the child node in the in-cluster tree aggregation network sends the model parameter to the parent node, and performs model parameter aggregation with the local model parameter of the parent node; Aggregate the intra-cluster model parameter aggregation results of each of the data homogeneity clusters to obtain a second-layer model parameter aggregation result, and when a global model using the second-layer model parameter aggregation result as a model parameter converges, send the global model to each of the edge computing devices. 14. An image processing device under data heterogeneity conditions, characterized by comprising: A partitioning module, configured to partition the edge computing devices into a plurality of data homogeneity clusters according to the similarity of data distribution of the edge computing devices; A selection module, configured to select an edge computing device in the data homogeneity cluster that is closest to other edge computing devices in the cluster or has the highest communication rate with other edge computing devices in the cluster as a cluster head of the data homogeneity cluster; A receiving module is used to receive the intra-cluster model parameter aggregation result of the data homogeneity cluster uploaded by the cluster head; the intra-cluster model parameter aggregation result is the first-layer model parameter aggregation result obtained by each edge computing device in the data homogeneity cluster aggregating local model parameters according to the intra-cluster tree aggregation network, each edge computing device in the data homogeneity cluster constructs the intra-cluster tree aggregation network according to the optimal communication strategy, the cluster head is the root node of the intra-cluster tree aggregation network, and the child nodes in the intra-cluster tree aggregation network send model parameters to the corresponding parent nodes, and perform model parameter aggregation with the local model parameters of the parent nodes; An aggregation module is used to aggregate the intra-cluster model parameter aggregation results of each of the data homogeneity clusters to obtain a second-layer model parameter aggregation result, and when the image processing model using the second-layer model parameter aggregation result as the model parameter converges, the image processing model is sent to each of the edge computing devices, so that the edge computing device uses the image processing model to process the image. 15. A federated learning device under heterogeneous data conditions, characterized by comprising: A division unit, configured to divide the edge computing devices into a plurality of data homogeneity clusters according to the data distribution similarity of the edge computing devices; A selection unit, configured to select an edge computing device from the data homogeneity cluster as a cluster head of the data homogeneity cluster; A receiving unit is used to receive the intra-cluster model parameter aggregation result of the data homogeneity cluster uploaded by the cluster head; the intra-cluster model parameter aggregation result is a first-layer model parameter aggregation result obtained by each edge computing device in the data homogeneity cluster aggregating local model parameters according to the intra-cluster tree aggregation network, each edge computing device in the data homogeneity cluster constructs the intra-cluster tree aggregation network according to the optimal communication strategy, the cluster head is the root node of the intra-cluster tree aggregation network, and the child nodes in the intra-cluster tree aggregation network send model parameters to the corresponding parent nodes, and perform model parameter aggregation with the local model parameters of the parent nodes; An aggregation unit is used to aggregate the intra-cluster model parameter aggregation results of each of the data homogeneity clusters to obtain a second-layer model parameter aggregation result, and when a global model using the second-layer model parameter aggregation result as a model parameter converges, the global model is sent to each of the edge computing devices. 16. A device, comprising: Memory for storing computer programs; A processor, used for implementing the steps of the image processing method under data heterogeneity conditions as described in any one of claims 1 to 12 or the steps of the federated learning method under data heterogeneity conditions as described in claim 13 when executing the computer program. 17. An image processing system under heterogeneous data conditions, characterized by comprising: Edge computing devices are used to train image processing models using local data, obtain local model parameters, and process images using image processing modules sent by edge cloud servers; The edge cloud server is used to divide the edge computing devices into several data homogeneity clusters according to the data distribution similarity of the edge computing devices; select the edge computing device with the shortest distance to other edge computing devices in the cluster or the highest communication rate with other edge computing devices in the cluster as the cluster head of the data homogeneity cluster; receive the cluster model parameter aggregation result of the data homogeneity cluster uploaded by the cluster head; the cluster model parameter aggregation result is the first layer model parameter aggregation result obtained by each edge computing device in the data homogeneity cluster according to the cluster tree aggregation network aggregation local model parameters, and each edge computing device in the data homogeneity cluster The edge computing device constructs the intra-cluster tree aggregation network according to the optimal communication strategy, the cluster head is the root node of the intra-cluster tree aggregation network, and the child nodes in the intra-cluster tree aggregation network send model parameters to the corresponding parent nodes, and perform model parameter aggregation with the local model parameters of the parent nodes; aggregate the intra-cluster model parameter aggregation results of each of the data homogeneity clusters to obtain the second-layer model parameter aggregation results, and when the image processing model using the second-layer model parameter aggregation results as the model parameters converges, the image processing model is sent to each of the edge computing devices, so that the edge computing devices use the image processing model to process the image. 18. A federated learning system under heterogeneous data conditions, characterized by comprising: Edge computing devices are used to train models using local data and obtain local model parameters; The edge cloud server is used to divide the edge computing devices into several data homogeneity clusters according to the data distribution similarity of the edge computing devices; select an edge computing device from the data homogeneity cluster as the cluster head of the data homogeneity cluster; receive the cluster model parameter aggregation result of the data homogeneity cluster uploaded by the cluster head; the cluster model parameter aggregation result is the first layer model parameter aggregation result obtained by each edge computing device in the data homogeneity cluster aggregating local model parameters according to the cluster tree aggregation network, each edge computing device in the data homogeneity cluster constructs the cluster tree aggregation network according to the optimal communication strategy, the cluster head is the root node of the cluster tree aggregation network, the child nodes in the cluster tree aggregation network send model parameters to the corresponding parent nodes, and perform model parameter aggregation with the local model parameters of the parent nodes; aggregate the cluster model parameter aggregation results of each data homogeneity cluster to obtain the second layer model parameter aggregation result, and when the global model with the second layer model parameter aggregation result as the model parameter converges, the global model is sent to each edge computing device. 19. A medium, characterized in that a computer program is stored on the medium, and when the computer program is executed by a processor, the steps of the image processing method under data heterogeneity conditions as described in any one of claims 1 to 12 or the federated learning method under data heterogeneity conditions as described in claim 13 are implemented.