CN119939637A

CN119939637A - Server cluster security control method combined with multi-level encryption strategy joint scheduling

Info

Publication number: CN119939637A
Application number: CN202510423341.8A
Authority: CN
Inventors: 张志杰; 刘佳; 赵帅
Original assignee: Beijing Hanxinsheng Technology Co ltd
Current assignee: Beijing Hanxinsheng Technology Co ltd
Priority date: 2025-04-07
Filing date: 2025-04-07
Publication date: 2025-05-06
Anticipated expiration: 2045-04-07
Also published as: CN119939637B

Abstract

The invention relates to the technical field of computers and discloses a server cluster security control method combining multi-level encryption strategy joint scheduling, which comprises the following steps of S1, performing sensitivity evaluation on calculation tasks and data in a cluster, classifying the data into low-sensitivity data, medium-sensitivity data and high-sensitivity data according to a preset rule, S2, encrypting the low-sensitivity data by adopting a first encryption algorithm, encrypting the medium-sensitivity data by adopting a second encryption algorithm, and encrypting the high-sensitivity data by adopting a homomorphic encryption algorithm, and S3, acquiring load, calculation task requirements and data sensitivity information of the cluster in real time by adopting a joint scheduling mechanism based on reinforcement learning. Through multistage encryption strategy, dynamic resource scheduling based on deep Q learning, and a security protection mechanism combining zero trust architecture and deep learning, the problems of low cluster resource utilization efficiency, insufficient data security and incapability of responding to security threats in real time are solved.

Description

Server cluster security control method combining multi-stage encryption strategy joint scheduling

Technical Field

The invention relates to the technical field of computers, in particular to a server cluster security control method combining multi-level encryption strategy joint scheduling.

Background

With the wide application of server clusters in the fields of big data, cloud computing and the like, efficient management of cluster resources and data security protection have become key problems. In conventional cluster management schemes, resource scheduling often relies on static configurations and rules, and cannot be dynamically adjusted according to cluster load changes, task demands, and data sensitivity. The static resource scheduling mode cannot effectively utilize cluster resources, and is easy to cause overload of computing nodes or idle resources, so that the overall performance of the cluster is affected.

In terms of data security, the existing encryption technology generally adopts a unified encryption strategy, and has no flexible encryption scheme for data with different sensitivities, so that the performance cost of a computing task is high. In addition, conventional encryption methods often have performance bottlenecks in large-scale clusters, and particularly, when processing large amounts of data, performance loss is more obvious.

Aiming at the security problem in the cluster, the traditional protection means generally depend on a trust-based model, neglect fine-grained access control and dynamic rights management, and cannot effectively prevent unauthorized access and potential internal threats. In addition, the variety of security threats and complex attack patterns present in clustered environments present significant challenges to traditional security safeguards, and a need for more intelligent and efficient security control strategies exists.

Disclosure of Invention

Aiming at the defects of the prior art, the invention provides a server cluster security control method combined with multi-level encryption strategy joint scheduling, which solves the problems of low resource scheduling efficiency, insufficient data security protection and insufficient capability of coping with dynamic security threat in a server cluster.

The server cluster security control method combining the multi-level encryption strategy joint scheduling comprises the following steps:

s1, performing sensitivity evaluation on computing tasks and data in a cluster, and classifying the data into low-sensitivity data, medium-sensitivity data and high-sensitivity data according to a preset rule;

s2, encrypting the low-sensitivity data by adopting a first encryption algorithm, encrypting the middle-sensitivity data by adopting a second encryption algorithm, and encrypting the high-sensitivity data by adopting a homomorphic encryption algorithm;

s3, a joint scheduling mechanism based on reinforcement learning dynamically adjusts the allocation of computing resources and task scheduling by acquiring the load, the computing task demand and the data sensitivity information of the cluster in real time;

S4, deploying a zero trust architecture in the cluster, enabling each request to be subjected to identity verification and authority authorization, and carrying out fine-grained control on task authorities through an access control strategy based on roles;

s5, monitoring calculation tasks and data transmission in the cluster in real time, performing anomaly detection on the cluster behaviors by adopting a deep learning algorithm, and identifying potential security threats;

And S6, when the potential threat is detected, automatically executing a safety response mechanism, and protecting by isolating abnormal nodes and limiting the resource access mode of the malicious task.

Preferably, the sensitivity evaluation of the data in the cluster in the step S1 specifically includes the following steps:

s1.1, collecting data of computing tasks in a cluster, and analyzing data types, contents and access frequencies of the data types and the contents;

S1.2, performing sensitivity evaluation on the data through a trained machine learning model or a specified statistical method, and dividing the data into low-sensitivity data, medium-sensitivity data and high-sensitivity data according to an evaluation result;

S1.3, classifying the data according to the evaluation result, and designating a corresponding encryption algorithm and a security policy according to the classification;

s1.4, updating the sensitivity level of the data periodically or in real time according to the change condition of the cluster resources, and dynamically adjusting the encryption strategy.

Preferably, the step S2 specifically includes the following steps:

S2.1, encrypting the low sensitive data by applying an AES-128 encryption algorithm to generate a corresponding encrypted ciphertext;

S2.2, encrypting the middle sensitive data by applying an AES-256 encryption algorithm to generate a corresponding encrypted ciphertext;

s2.3, encrypting the high-sensitivity data by applying a homomorphic encryption algorithm, wherein the homomorphic encryption algorithm is a Paillier encryption algorithm;

s2.4, all encryption processes generate and manage keys through a centralized key management system, wherein the key management system comprises life cycle management, key storage and distribution.

Preferably, the joint scheduling mechanism based on reinforcement learning in the step S3 includes:

s3.1, collecting the resource use condition of the current cluster, including the load, the memory and the CPU use rate of the computing node, and the computing requirement and the priority of the task;

S3.2, dynamically adjusting a task scheduling strategy and a resource allocation strategy under the condition that the resource allocation constraint condition is met based on deep Q learning;

S3.3, selecting a computing node which is free in resources and can meet encryption computing requirements according to the encryption level, data sensitivity and computing resource requirements of the task;

and S3.4, in the scheduling process of the calculation task, the resource utilization rate, load balancing and safety are considered.

Preferably, the implementation of the zero trust architecture in step S4 specifically includes:

s4.1, carrying out OAuth2.0 or OpenIDConnect-based identity verification on each request in the cluster;

S4.2, realizing access control based on attributes, and controlling access rights according to the sensitivity of the task, the identity and role information of the requester;

s4.3, checking authority after each request access;

s4.4, dynamically adjusting the permission, and enabling the permission to be updated in real time in the life cycle of the task according to the task requirements and the access condition.

Preferably, the step S5 specifically includes:

s5.1, monitoring the use condition of the computing node resources of the cluster in real time by using a Prometaus monitoring system, and visually displaying through Grafana;

s5.2, modeling the cluster behaviors by using a deep learning algorithm, and identifying an abnormal mode deviating from the normal behaviors;

s5.3, automatically triggering an alarm and notifying an administrator when the cluster resource or task behavior is abnormal by using an abnormality detection method based on a model;

S5.4, combining the monitoring system with the abnormal detection result to generate a real-time safety report so as to help an administrator to take countermeasures in time.

Preferably, the step S6 secure response mechanism includes:

s6.1, when the security threat is detected, automatically isolating a malicious node, stopping a computing task on the node, and preventing the security threat from spreading;

s6.2, limiting the resource access rights of the affected nodes or tasks, and reducing the interference of the affected nodes or tasks to normal computing tasks;

s6.3, according to the severity of the abnormal behavior, automatically adjusting the access authority of the cluster, and prohibiting unauthorized users or nodes from accessing sensitive data;

S6.4, recording all security events in detail, generating an audit log, and providing the audit log for an administrator to carry out subsequent analysis and investigation.

Preferably, the joint scheduling mechanism of reinforcement learning in the step S3 optimizes the scheduling policy by the following mathematical model: Wherein R (t) represents a reward value of a scheduling policy, R _i represents the available capacity of a resource i, t _j represents the demand of a task j, x _ij is a Boolean variable, whether the task j is allocated to the resource i, alpha and beta are regularization coefficients, n is the number of resource nodes, m is the number of tasks, d _j is the resource demand of the task j, and p _j is the priority of the task j.

The invention provides a server cluster security control method combined with multi-level encryption strategy joint scheduling. The beneficial effects are as follows:

1. The invention realizes the encryption protection of the data with different sensitivity levels by carrying out sensitivity evaluation on the calculation tasks and the data in the clusters and selecting a proper encryption algorithm according to the evaluation result, compared with the traditional unified encryption scheme, the multi-stage encryption strategy of the invention can carry out customized encryption according to the specific sensitivity of the data, the flexibility and the safety of data encryption are greatly improved, the encryption burden of low-sensitivity data can be effectively reduced through the flexible encryption management, the high-sensitivity data is ensured to be protected more strongly, and therefore the overall safety of the cluster is ensured, and the resource utilization efficiency is improved.

2. The invention can acquire cluster load, task demand and data sensitivity information in real time and dynamically adjust the distribution of computing resources and task scheduling by introducing a reinforcement learning algorithm, in particular to a joint scheduling mechanism based on deep Q learning. Compared with the traditional static resource allocation or simple scheduling algorithm, the dynamic scheduling mechanism based on reinforcement learning can optimize according to the workload and task demands which change in real time, and automatically select the most suitable computing resource for task scheduling, so that the utilization efficiency and task processing capacity of cluster resources are improved.

3. The invention combines the zero trust architecture and the deep learning algorithm, realizes strict access control and real-time abnormal behavior detection in the cluster, all requests are required to be subjected to identity verification and authority authorization, and simultaneously, fine-granularity access control is carried out according to task sensitivity and user role information.

Drawings

FIG. 1 is a flow chart of the present invention.

Detailed Description

The following description of the embodiments of the present invention will be made clearly and fully with reference to the accompanying drawings, in which it is evident that the embodiments described are only some, but not all embodiments of the invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

Referring to fig. 1, an embodiment of the present invention provides a server cluster security control method combined with multi-level encryption policy joint scheduling, including the following steps:

In the step S1, the sensitivity evaluation of the data in the cluster specifically comprises the following steps:

Specifically, in step S1.1, the system first needs to collect all relevant computing task data from each computing node in the cluster. The data collection includes not only the type and content of the data, but also the access frequency, access pattern and historical usage of the data. For example, some data may be frequently accessed by different computing nodes, while other data may be used only in certain tasks. Therefore, the access frequency and pattern are one of the important bases for sensitivity evaluation. The information can be obtained in real time through a cluster management system or a special log analysis system, so that the comprehensiveness and accuracy of data collection are ensured.

In step S1.2, based on the collected task data, the system evaluates the sensitivity of the data using trained machine learning (e.g., support vector machine, decision tree, random forest, etc.) or statistical methods. The common sensitivity assessment model can comprehensively obtain the sensitivity level of each data based on the type of the data (such as financial data and user information), content (such as sensitive fields and privacy data), access frequency and other characteristics. The model in the data evaluation process can be trained through historical data and is continuously adjusted and optimized so as to improve the evaluation accuracy.

The output results of the sensitivity assessment divide the data into three categories:

Low sensitivity data, which may be public data or data with low security requirements, may be protected using conventional encryption or simple access control measures.

Medium sensitive data, which need moderate protection measures, are processed by adopting a strong encryption algorithm and proper access control.

Highly sensitive data, which are extremely sensitive core data, require the highest level of protection, are typically encrypted using advanced encryption techniques such as homomorphic encryption, and have their access rights tightly controlled. The formula is to set the data set asWherein d _i is the ith data, data sensitivity assessment is performed by the following model: Wherein S (d _i) represents the sensitivity evaluation result of the data d _i, f is a trained sensitivity evaluation model, and is output as the sensitivity level of the data.

Based on the evaluation results, the system classifies the data according to its sensitivity level. Specifically, for different classes of data, the system specifies different encryption algorithms and access control policies:

For low-sensitivity data, an AES-128 and other simpler encryption algorithms are adopted, and the low-sensitivity data is protected through basic identity authentication and access control;

for medium sensitive data, an AES-256 encryption algorithm is adopted for protection, and a stricter access right is configured for the data;

For highly sensitive data, homomorphic encryption algorithms (e.g., paillier encryption algorithms) are used to encrypt data to ensure that the data can be processed and calculated even in the encrypted state.

In this step, the system corresponds the encryption algorithm to the sensitivity level of the data one by one, and ensures that each data class has corresponding protection measures. Meanwhile, the system can allocate proper computing resources and scheduling strategies for the data of different levels according to task requirements, and ensure that encryption computation cannot cause excessive influence on cluster performance.

As the computing tasks in the cluster change and the data access patterns adjust, the sensitivity of the data may change. Thus, in this step, the system needs to update the sensitivity level of the data periodically or according to real-time conditions. This is accomplished by continuous monitoring and data evaluation, ensuring that the data is always at the proper level of security protection throughout the life cycle.

In the updating process, the system dynamically adjusts the data encryption strategy according to the new sensitivity evaluation result. For example, when the access frequency of some sensitive data increases significantly, the system may upgrade it to highly sensitive data and switch to a stronger encryption algorithm. Meanwhile, the system can optimize the resource allocation of encryption calculation according to the priority of the task and the change of the calculation resource, and ensure the balance between data protection and cluster resources.

According to the embodiment, sensitivity evaluation is carried out on the data by combining a machine learning method and statistical analysis, and a multi-level encryption strategy is implemented according to an evaluation result, so that the data in the cluster can be properly protected according to different security requirements. In the sensitivity evaluation process, the system combines various factors such as task types, data access frequency and the like, and the accuracy of sensitivity division is ensured by utilizing an intelligent evaluation model. The method has higher flexibility and expandability, and can effectively cope with various security threats in a cluster environment.

The step S2 specifically comprises the following steps:

Specifically, in this embodiment, the step S2 encrypts the data, which specifically includes the following contents:

For low sensitivity data, the AES-128 (advanced encryption standard, 128 bit key length) encryption algorithm is used. AES-128 is a symmetric encryption algorithm, has relatively simple encryption process and high calculation speed, and is suitable for scenes with large data volume and relatively low security requirements. Although the data encrypted by the AES-128 algorithm has certain protection capability, the encryption strength is lower, so that the method is suitable for protecting data which is not sensitive or not easy to abuse, such as some disclosed system logs or other data with lower requirements on security.

In this step, the system encrypts the low sensitivity data using a 128 bit key, generating a corresponding encrypted ciphertext. The encrypted data will be stored or transmitted to the computing nodes of the cluster, and during transmission, the data will be protected from unauthorized access.

The sensitive data in the medium often includes information important in some businesses (such as user data, some business data, etc.), so stronger encryption measures are required. Compared with AES-128, the AES-256 uses a 256-bit key for encryption, so that stronger safety guarantee is provided, and attacks such as brute force cracking can be effectively prevented.

In this step, the system generates a 256-bit key for the medium sensitive data and encrypts the data using the AES-256 algorithm. The data ciphertext encrypted by the AES-256 is safer, can bear stronger attack, and is suitable for protecting data with certain confidentiality but not extremely sensitive.

Highly sensitive data is the most confidential, most protected type of data in a cluster, such as financial data, personal privacy data, core business data, etc. In order to ensure that the data can still be calculated and analyzed after being encrypted, the homomorphic encryption algorithm, particularly the Paillier encryption algorithm, is selected in the embodiment.

Homomorphic encryption is a special encryption scheme that allows certain computations to be performed directly on encrypted data without decryption, thereby avoiding the risk of exposing the data during processing. The Paillier encryption algorithm is an addition homomorphic encryption algorithm, and is particularly suitable for a scene requiring addition operation on encrypted data, such as summation operation on encrypted financial data without decryption. The Paillier algorithm encrypts and decrypts data by generating a public key that is used to encrypt the data and a private key that is used to decrypt the data.

By the encryption algorithm, highly sensitive data can be calculated in an encrypted state, so that the security and the accuracy of a calculation result are ensured while data leakage is prevented.

The management of keys is a core part of the encryption system. In this step, all encryption processes rely on a centralized key management system to generate, distribute, and store keys. The key management system not only generates keys (such as AES-128, AES-256 and Paillier keys) required by the encryption algorithm, but also is responsible for managing the life cycle of the keys, including operations such as key creation, updating, revocation, storage and distribution.

In particular, key management systems need to ensure secure storage of keys, preventing access or disclosure of keys by unauthorized users. For this reason, keys are typically stored in a Hardware Security Module (HSM) or key management hardware, providing a high level of protection. The key management system also needs to ensure that the distribution process of the key is secure, and only nodes that pass authentication and authorization can acquire the key.

In addition, the key management system will update the key periodically to reduce the risk of the key being hacked. For expired or no longer needed keys, the system can conduct revocation operation, and the validity and the security of the keys are ensured.

In the embodiment, AES-128, AES-256 and Paillier homomorphic encryption algorithms are respectively applied to low-sensitive data, medium-sensitive data and high-sensitive data in a cluster through a multi-level encryption strategy. Different encryption algorithms ensure the security of data with different sensitivity levels, and simultaneously reduce the consumption of computing resources to the greatest extent. All encryption processes carry out key generation, storage, distribution and life cycle management through a centralized key management system, so that the security and effectiveness of the keys are ensured. The scheme provides a flexible and efficient solution for data security in the cluster environment, and is beneficial to optimizing cluster resource utilization efficiency while ensuring security.

The joint scheduling mechanism based on reinforcement learning in the step S3 comprises the following steps:

And S3, optimizing a scheduling strategy by a reinforcement learning joint scheduling mechanism through the following mathematical model: Wherein R (t) represents a reward value of a scheduling policy, R _i represents the available capacity of a resource i, t _j represents the demand of a task j, x _ij is a Boolean variable, whether the task j is allocated to the resource i, alpha and beta are regularization coefficients, n is the number of resource nodes, m is the number of tasks, d _j is the resource demand of the task j, and p _j is the priority of the task j.

Specifically, in this embodiment, the resource usage of the cluster includes the load of each computing node, the memory and the CPU usage, and the computing requirement and priority information of the current computing task. Specifically:

And (3) collecting the load of the computing nodes, namely monitoring the resource use condition of each node in real time by a cluster management system, wherein the load state of the computing node i is measured by two indexes, namely cpu _i and mem _i. The load function of a node can be expressed as: Wherein f is a load function, which represents that the load of the node is comprehensively judged through the CPU and the memory resource consumption of the computing node.

Task computing requirements and priorities, each computing task j having its computing resource requirementsAnd wherein: And The demands of the task j on the CPU and the memory resources are respectively met, p _j is the priority of the task, and the task with higher priority can be scheduled preferentially.

By the method, the system can accurately collect the resource use condition of all nodes in the cluster and the demand state of the task, and data support is provided for subsequent scheduling decisions.

After the cluster resources and task information are collected, the system then uses the deep Q learning method in reinforcement learning to implement dynamic task scheduling and resource allocation. The specific steps are as follows:

Definition of state space:

the state space of the system contains the resource states of all the computing nodes and the demand information of the tasks to be scheduled. The state s _t may be represented as a set of resource conditions and current tasks to be scheduled for all nodes in the cluster: Wherein, the method comprises the steps of, Representing the resource status of the ith node,And represents the resource requirements and priorities of the jth task.

Definition of action space a _t represents the selection of the current scheduling decision. For each task j, the system may choose whether it is assigned to node i action x _ij is a boolean variable that indicates whether task j is assigned to resource node i: The design of the reward function aims at maximizing the utilization efficiency of cluster resources, optimizing load balancing and considering the priority of tasks. Based on the assignment of tasks, the reward function R (t) can be defined by the following formula: Wherein R (t) represents a reward value of a scheduling policy, R _i represents the available capacity of a resource i, t _j represents the demand of a task j, x _ij is a Boolean variable, whether the task j is allocated to the resource i, alpha and beta are regularization coefficients, n is the number of resource nodes, m is the number of tasks, d _j is the resource demand of the task j, and p _j is the priority of the task j.

This reward function aims to optimize the scheduling of resources by taking into account the utilization of the resources and the priority of the tasks so that the tasks can be efficiently scheduled to the appropriate computing nodes.

Q value update rule in reinforcement learning, Q value represents the expected return for taking an action in a certain state. After each scheduling decision, the system updates the Q value based on the rewards obtained. The update of the Q value follows the classical Q learning formula: Wherein α is the learning rate, γ is the discount factor, and r _t is the prize obtained at the current time step. And is the maximum Q value for all possible actions in the next state s _t+1.

In this embodiment, the task is scheduled taking into account not only its computing resource requirements and priorities, but also the encryption level and data sensitivity of the task. According to different encryption requirements of the tasks, the system dispatches the tasks to the proper computing nodes so as to meet the security and computing requirements:

Encryption level-low sensitive tasks use standard encryption (e.g., AES-128), medium sensitive tasks use stronger encryption (e.g., AES-256), and high sensitive tasks use homomorphic encryption algorithms (e.g., paillier encryption).

And (3) the demand of encryption computing resources, namely, as the encryption algorithm consumes different computing resources, the system dynamically calculates the resource demand according to the encryption level of the task. Encryption computing requirementsThe calculation can be performed by the following model: Wherein f _enc is an encryption requirement function, which represents the resource consumption of tasks under different encryption levels.

And (3) resource allocation, namely selecting a computing node capable of providing enough computing capacity by the system based on the encryption requirement of the task, and ensuring that the encryption task can be successfully completed.

In the scheduling process, the system comprehensively considers the following factors:

Resource utilization and load balancing to avoid overload of certain nodes in the cluster, the system may make a tradeoff between load balancing and resource utilization. For example, when a node is more than 80% loaded, the system may avoid assigning new tasks to it.

And for the high-sensitivity task, the system preferably selects the node with reinforced safety protection for scheduling so as to ensure that the transmission and calculation process of the task data meet the safety requirements.

the implementation of the zero trust architecture in the step S4 specifically includes:

s4.3, checking authority after each request access;

Specifically, in this embodiment, in order to ensure that each request is subjected to strict authentication, authentication is performed in the cluster using an authentication protocol based on oauth2.0 or OpenIDConnect. These protocols can effectively implement validation of the user identity and ensure validity of the requester identity.

Oauth2.0 authentication oauth2.0 protocol allows client applications to access a resource server on behalf of a user. The client must acquire an authorization token as the requested credential through the authorization server. Each time a cluster resource is requested, the client needs to carry an access token of oauth2.0, and the server verifies the identity to the authorization server through the token.

OpenIDConnect authentication OpenIDConnect is an extension of oauth2.0 protocol, providing additional authentication functionality. Through OpenIDConnect, the cluster can verify the identity of the requesting party and obtain the user's basic information (e.g., name, email, role, etc.). Each request is authenticated by the identity provider and an ID token is generated for subsequent authentication of rights.

The authentication process is that the requester first requests authentication from the authentication server through OAuth2.0 or OpenIDConnect protocol and obtains the identity token.

The token is sent with the request to the cluster server, which validates the identity of the requestor by verifying the validity of the token.

If the authentication is successful, the subsequent authorization stage is entered, and if the authentication is failed, the request is refused and the corresponding error information is returned.

Attribute-based access control is a more flexible and dynamic rights control mechanism that determines access rights by evaluating multiple attributes of a requestor (e.g., identity, role, task sensitivity, etc.). Access control can provide more sophisticated rights management than traditional role-based access control.

Access control policies in attribute-based access control, the access control policies rely on a set of defined attributes, which can be divided into the following categories:

requester attributes including identity information, roles, groups, etc. of the user.

Task attributes including sensitivity level of the task, lifecycle status of the task, etc.

Environmental attributes including time of request, geographic location, source of request, etc.

Policy example:

For example, a highly sensitive task (task sensitivity high) may be accessible only by a particular group of users (e.g., a system administrator group) and may be accessible only during a working time (e.g., 9:00-18:00). The policy may be defined as: By such policy definition, attribute-based access control can ensure that rights decisions are made dynamically and flexibly in the cluster according to task and requestor specifics.

In the implementation mode, an access control system in the cluster matches the request with related attributes and judges whether access is allowed or not according to a preset access control strategy.

If the requester satisfies all policy conditions, the system will allow the request to be executed, otherwise, a response is returned to deny access.

Under a zero trust architecture, each request must go through a permission check, whether the source of the request is legitimate or from a known user. All requested access rights are temporary, dynamic, so each access triggers rights verification.

Rights verification process whenever a cluster receives a request, the system performs rights verification based on the context of the current request (including the identity of the requestor, the content of the request, the sensitivity of the task, etc.).

Rights verification generally includes the following aspects:

Authentication, checking whether the supplicant successfully authenticates through oauth2.0 or OpenIDConnect protocols.

Role and attribute matching by checking whether the requester has sufficient rights to access the target resource through attribute-based access control or a role-based access control model.

Task sensitivity and access control, further judging whether the access condition is met according to the sensitivity of the task and the attribute of the requester.

Verification failure if the rights verification fails, the system immediately terminates execution of the request and returns a corresponding error code (e.g., "403 Forbidden") to the requestor.

In order to better adapt to the change of the life cycle of the task and the difference of access modes, the authority needs to be adjusted in real time according to the requirements of the task and the access conditions. The system can dynamically update the authority to ensure the security of the task at different stages.

And the triggering condition of the permission adjustment is that when the sensitivity of the task changes, the system automatically adjusts the access permission. For example, when a task enters a high sensitivity phase, the system may increase the stringency of the rights, limiting access to users of a particular role only.

When the access frequency of the task increases, the system may temporarily adjust the resources and the permissions according to the load condition, so as to prevent potential safety hazards caused by excessive authorization or access to excessive resources.

Based on the life cycle of the task and the access log, the cluster management system dynamically updates the access rights according to preset rules. For example, if the task enters the sensitive phase, the permission check level is automatically raised.

In addition, the cluster will log each entitlement change for auditing and tracking purposes. Each authority adjustment operation can generate an audit log and synchronize to the management platform in real time for monitoring and management.

The step S5 specifically comprises the following steps:

Specifically, in order to realize real-time monitoring of cluster computing tasks and resources, in this embodiment, prometaus is used as a cluster resource monitoring system, and the Prometaus can regularly pull resource usage indexes of nodes of a cluster and store the resource usage indexes as time sequence data. Grafana is used for visually displaying the Prometheus monitoring data so that an administrator can intuitively know the health state of the cluster.

Prometaus monitoring system Prometaus is an open source monitoring system, can periodically collect resource use condition of each node, and supports efficient storage, inquiry and alarm. It mainly monitors the following aspects:

And calculating node resource usage such as CPU usage, memory usage, disk IO, network bandwidth and the like.

Task execution conditions such as running time, state, resource requirement and the like of the task.

Prometaus collects and saves the resource information of each computing node through exporters to generate time sequence data.

Grafana A visualization Grafana is used as a visual display tool of Prometheus, and the real-time collected resource usage data is presented to an administrator in the form of a chart, a dashboard and the like. Through Grafana, an administrator can see the resource usage of each node in real time to find potential performance bottlenecks or anomalies.

The implementation flow is as follows:

deployment Prometheus installs a monitoring agent or exporter on each compute node, configures Prometheus timing pull data.

The configuration Grafana interfaces with Prometheus, and the visual dashboard is configured through Grafana to show the running status, resource usage, task execution, etc. of the system.

And a monitoring report is generated regularly, so that an administrator can conveniently conduct performance analysis.

In order to more efficiently identify abnormal behaviors in a cluster, a system adopts a deep learning algorithm to perform cluster behavior modeling, identifies a mode deviating from normal behaviors, and further detects potential security threats.

And constructing a deep learning model, namely taking resource monitoring data, task execution data, network traffic and the like collected in the cluster as input, and training the model by using a deep learning algorithm. These models are able to learn normal operation modes and behavior modes.

Self-encoder-by self-encoder model, the normal mode of cluster behavior is measured by the reconstruction loss of the inputs and outputs. When abnormal behavior occurs, the reconstruction error may increase significantly and thus be used to detect behavioral anomalies.

Firstly, collecting a large amount of cluster data which normally run, including the use condition of computing resources, the task state, the data transmission condition and the like.

These data are trained using a deep learning algorithm (e.g., LSTM model) to obtain a "normal behavior model" of the cluster.

In actual operation, real-time monitoring data are input into a deep learning model, and whether abnormality exists or not is judged by calculating a loss value.

And (3) abnormal pattern recognition, namely when the behavior of the real-time monitoring data deviates from the normal behavior learned by the model, the model outputs a larger abnormal score to indicate that the current cluster is abnormal.

Such anomalies include, but are not limited to, compute node load anomalies, network traffic surges, task execution time overlengths, data transmission speed anomalies, and the like.

The abnormality detection method based on the model is to judge whether abnormality occurs by comparing the difference between the cluster behavior and the trained normal behavior model. When an abnormality is found, the system can automatically trigger an alarm mechanism to inform an administrator to take countermeasures in time.

The system judges whether the abnormality exists or not by comparing the real-time monitoring data with the normal mode in the deep learning model. If the output of the model prediction is greatly different from the actual monitoring data, the cluster behavior is deviated from the normal mode, and the cluster behavior is judged to be abnormal.

And alarming based on a threshold value, namely triggering an alarm when the abnormal score of a certain resource exceeds a set threshold value.

And the automatic alarm mechanism is that the system can automatically trigger an alarm according to a preset strategy and a threshold value. These alarms include:

and alarming system performance, such as overhigh CPU utilization rate, abnormal memory occupation and the like.

Task execution abnormality alarms such as task running overtime, resource request overhigh and the like.

Network abnormality alarms such as network traffic bursts, abnormal data transmission behavior, etc.

Notification mechanism-alarm information is pushed to the administrator in real time, and notification content includes anomaly type, influence range, specific index and possible solution.

The administrator may be notified through various channels, such as email, text messaging, instant messaging, etc.

The generation of the real-time security report is the final output of cluster monitoring and anomaly detection, helping the administrator to know the health status of the current cluster and respond quickly to possible security events.

And generating a report, namely acquiring the resource use condition of each node in real time by a monitoring system in the cluster operation process, and generating a detailed safety report by combining the deep learning abnormal detection result. Report content typically includes:

and the current resource use condition is the use condition of CPU, memory and network bandwidth of each node.

And (3) detecting the abnormality, namely judging whether abnormal behaviors exist or not, and judging the type and the range of the abnormality.

Alarm record, namely detail record of abnormal events occurring in the system.

Report format the real-time security report can be presented in various forms such as charts, logs, warnings, etc., and the report is sent to an administrator or security operation team immediately after being generated.

Report content includes analysis of abnormal behavior, potential risk assessment, and recommended countermeasures such as limiting access to abnormal nodes, optimizing resource allocation, etc.

The report helps the manager to quickly know the health status of the cluster, and timely discover and deal with potential security threats.

For example, if security risks are caused by network traffic anomalies, an administrator may immediately take the option of quarantining the relevant nodes or restricting access to certain services.

The S6 step safety response mechanism comprises the following steps:

In particular, when a system detects a malicious node or potential security threat, quarantine measures must be quickly taken to prevent the threat from spreading to other computing nodes or affecting other parts of the system.

And detecting malicious nodes, namely judging whether abnormal behaviors of the nodes occur or not by the system through a deep learning model, a behavior analysis algorithm and real-time monitoring data, such as task timeout, abnormal resource use, abnormal network flow and the like.

If a significant departure of node behavior from normal mode is detected and characteristics of malicious activity (e.g., malicious scripts, etc.) are met, the system will determine that the node is a "malicious node".

And automatically isolating the malicious node, wherein once a certain node is determined to be the malicious node, the system immediately isolates the node. The isolation measures include:

Network isolation, namely disconnecting the malicious node from the cluster internal network and preventing the malicious node from communicating with other normal nodes.

And stopping tasks, namely forcibly stopping all the computing tasks on the node, and preventing further execution of malicious tasks.

And (3) resource recovery, namely immediately recovering the computing resources occupied by the node and releasing the computing resources to other nodes for normal task processing.

Threat spreading is prevented, namely, malicious behaviors can be prevented from spreading to other nodes in the cluster or affecting key data processing tasks by rapidly isolating the malicious nodes, so that the overall security of the cluster is ensured.

After the threat is detected, besides isolating the malicious node, the resource access authority of the affected node or task needs to be limited, so that the interference to the normal computing task is prevented.

Resource access restrictions for nodes or tasks that are determined to be affected, the system will take steps to restrict its resource access rights, including:

And (3) limiting the CPU/memory, namely limiting the quota of the CPU and the memory resources of the affected task, and preventing the CPU and the memory resources from occupying excessive resources.

Network access limitations-limiting the network bandwidth or network access capacity of the affected node, reducing its interference to other nodes.

And the storage access limitation is that the access of the affected node to the storage system is limited, and malicious data tampering or leakage is prevented.

The interference is reduced, namely, the resource use of the affected node or task is accurately controlled, so that the influence on other normally-running computing tasks is ensured to be minimized, and the potential safety risk can be effectively isolated.

According to the severity of the detected abnormal behavior, the system needs to dynamically adjust the access rights in the cluster, so as to ensure that sensitive data in the cluster is not accessed by unauthorized nodes or users.

Severity assessment of abnormal behavior:

the system will evaluate the severity of the detected abnormal behavior based on its nature and impact. For example, if a node presents malicious code or network attack behavior, the system may evaluate the behavior for a potential threat to the data.

Severity assessment criteria the severity of abnormal behavior can be assessed based on factors such as duration of abnormal behavior, scope of influence, task sensitivity, etc.

And (3) access right adjustment:

When the abnormal behavior is estimated to be serious, the system automatically adjusts the access right according to the security policy of the cluster:

And prohibiting unauthorized users from accessing the sensitive data, wherein if a certain node is determined to be a malicious node, all requests of the node for accessing the sensitive data are prohibited.

Enhanced authentication, namely, enhanced authentication is carried out on all users and nodes, so that only authorized nodes can access key resources and sensitive data.

Limiting data transmission authority, namely limiting data transmission among tasks and preventing sensitive data from being illegally accessed or leaked.

Dynamic adjustment, namely, adjustment of access authority can be performed in real time, and normal access authority strategies can be restored after abnormal behaviors are repaired or threatened is relieved.

During the process of executing the safety response mechanism, the system can record all the occurred safety events in detail, and generate an audit log so that an administrator can conduct subsequent analysis and investigation to ensure the transparency and traceability of the events.

Security event logging-all operations relating to security threat detection, quarantine, rights adjustment, etc. are logged. The recording content includes:

event time, the specific time when the security event occurs.

Event types such as "malicious node quarantine", "task stop", "rights adjustment", etc.

The scope of influence is the affected node, task or data.

Responsive measures, namely, safety responsive measures such as network isolation, resource limitation and the like are adopted.

And generating an audit log, namely storing the audit log in a safe log management system to ensure that the log is not tampered or deleted.

These log records can help administrators track and trace security events, further analyzing the source, process, and impact of malicious attacks.

The audit log is not only an important reference basis for safety response, but also provides data support for post analysis. The administrator may be based on log content:

and carrying out deep analysis of malicious behaviors and identifying an attack mode.

And evaluating the validity of the cluster security policy, and adjusting the security policy for the discovered loopholes.

Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents.

Claims

1. The server cluster security control method combining the multi-stage encryption strategy joint scheduling is characterized by comprising the following steps of:

2. The server cluster security control method combined with multi-level encryption policy joint scheduling according to claim 1, wherein the sensitivity evaluation of the data in the cluster in step S1 specifically includes the following steps:

3. The server cluster security control method combined with multi-level encryption policy joint scheduling according to claim 1, wherein the step S2 specifically includes the steps of:

4. The server cluster security control method in combination with multi-level encryption policy joint scheduling according to claim 1, wherein the joint scheduling mechanism based on reinforcement learning in step S3 includes:

5. The server cluster security control method combined with multi-level encryption policy joint scheduling according to claim 1, wherein the implementation of the zero trust architecture in step S4 specifically includes:

s4.3, checking authority after each request access;

6. The server cluster security control method combined with multi-level encryption policy joint scheduling according to claim 1, wherein the step S5 specifically includes:

7. The server cluster security control method combined with multi-level encryption policy joint scheduling according to claim 1, wherein the S6 step security response mechanism includes:

8. The server cluster security control method combined with multi-level encryption policy joint scheduling according to claim 1, wherein the joint scheduling mechanism of reinforcement learning in step S3 optimizes the scheduling policy by the following mathematical model: Wherein R (t) represents a reward value of a scheduling policy, R _i represents the available capacity of a resource i, t _j represents the demand of a task j, x _ij is a Boolean variable, whether the task j is allocated to the resource i, alpha and beta are regularization coefficients, n is the number of resource nodes, m is the number of tasks, d _j is the resource demand of the task j, and p _j is the priority of the task j.