CN109448744A - A kind of MP3 audio information hiding method and system based on sign bit adaptive feed-forward network - Google Patents

Abstract

The invention relates to an MP3 audio information hiding method and system based on symbol bit adaptive embedding. The steps of the method include: 1) extracting the sign bit of the qualified QMDCT coefficient in the large value area and the small value area of the MP3 audio as a carrier embedded in the secret information; 2) according to the modification range of the QMDCT coefficient and the psychoacoustic model The auditory threshold of the human ear calculates the modification cost of each embeddable position; 3) according to the modification cost of each embeddable position, the STC encoding is used to select the embedding path with the least distortion, and the embedding of the secret information is completed by modifying the sign bit of the QMDCT coefficient. The steganographic method of the invention has adaptability. By calculating the modification cost of each embedded point and combining STC coding, the embedded information is embedded at the embedding point with the least distortion cost, so that the concealment is high and the anti-detection property is high. And the steganographic capacity of the method of the invention is large.

Description

A kind of MP3 audio information hiding method and system based on sign bit adaptive feed-forward network

Technical field

The present invention relates to a kind of adaptive steganography methods and system for MP3 audio, belong to Information hiding field, especially It is related to hiding and extraction for secret information.

Background technique

Steganography be it is a kind of secret information is embedded into the multimedia files such as image and audio-video, with Communication hiding behavior Information Hiding Techniques, it can guarantee the transmission safe and out of sight of secret information.MP3 is the most popular audio of current internet One of format has the characteristics that file size is moderate, redundant space is big and coding is efficient, is a kind of good steganography carrier.

Currently, some Information Hiding Algorithms for MP3 compression domain have been proposed both at home and abroad.Wherein, MP3Stego is One of a most classical, most popular MP3 steganography software, by after MP3 cataloged procedure modification quantization step makes coding The parity of block length and information to be embedded, which are consistent, carrys out embedding information (bibliography: Petitcolas F.MP3Stego [EB/OL]).Feature when Yan Diqun is directed to Huffman code word in coding standard proposes a kind of by the secret of code word replacement realization Confidential information steganographic algorithm, the algorithm first classify to the code word in code table, to guarantee that replacement operation will not change code stream Fixed structure (bibliography: Yan D Q, Wang R D, Zhang L G.A high capacity MP3steganography based on Huffman coding[J].Journal of Sichuan University,2011,48(6):1281- 1286).It opens Yao et al. and proposes a kind of rear-mounted MP3 steganographic algorithm, algorithm code word lowest bit in the area little Zhi Hb code table Position insertion secret information, and can be according to the adaptive selection embedded length of secret information and carrier lengths (Yao, Pan Feng, Shen Inline type adaptive steganographic algorithm [J] computer engineering of the big of army based on MP3 and design, 2016,37 (6): 1537-1542). Dong Yakun proposes a kind of improvement LSB steganographic algorithm for Huffman code word linbits, and further utilizes decoding block class Type is selected (bibliography: the Beijing Investigation of Information Hiding Technology [D] of Dong Yakun based on MP3: Beijing postal to embedded location TV university, 2015).Yang Kun et al. proposes the adaptive steganographic algorithm of MP3 based on the replacement of isometric entropy coding, and the algorithm is right first Isometric code word carries out odd even distribution, using the bit vectors after distribution as carrier, then constructs cost using psychoacoustic model Function finally realizes adaptive steganography (bibliography: Yang K, Yi X W, Zhao X F, et using STC coding al.Adaptive MP3Steganography Using Equal Length Entropy Codes Substitution [C]//International Workshop on Digital Watermarking.Springer,LNCS 1043,2017: 202-216).Chinese invention patent CN103106901B proposes the MP3 steganography method based on audio particle index value.Middle promulgated by the State Council Bright patent CN102097098A proposes the compressed domain audio steganography method using window type switching law as foundation.Chinese invention Patent CN106228981A proposes a kind of adaptive steganographic algorithm of MP3 of directly modification frequency domain quantization parameter.

But the above-mentioned generally existing following shortcoming of steganography method: (1) safety is weaker.Using based on Markov spy Sign 80% or more can achieve to the Detection accuracy of above-mentioned non-adaptive steganography method (bibliography: Jin C, Wang R D, Yan D Q.Steganalysis of MP3Stego with low embedding-rate using Markov feature [J].Multimedia Tools&Applications,2016,76(5):1-16).(2) steganographic capacity is low.For example, The maximum steganographic capacity of MP3Stego algorithm only has 154bps or so (two-channel, 44.1KHz audio carrier).(3) Yang Kun et al. The adaptive steganographic algorithm of the MP3 of proposition needs to construct code word mapping space before steganography operation, increases computing cost.

Summary of the invention

For the deficiency for solving above-mentioned technology, the invention proposes one kind based on quantization MDCT (Quantized Modified Discrete Cosine Transform, QMDCT) coefficient symbols position modification the adaptive steganography method (Adaptive of MP3 Coefficient Sign-bit Steganography, ACS) and system.

It is found through principle analysis and experimental verification, modification QMDCT coefficient symbols position will not cause the change of frame length, can Compression flow structure is kept, and perceptual distortion caused by steganography can be controlled by the threshold value of limitation modification coefficient.Therefore, it selects QMDCT coefficient symbols position in threshold value is feasible as embedded domain.Before embedding information, with human ear in psychoacoustic model Muting threshold and modification amplitude are to calculate the modification cost of each embeddable point according to reasonable cost function is constructed.Information is embedding Enter to encode by STC and realize, in conjunction with the modification cost of each insertion point, optimal route selection is realized, so that insertion total distortion is most It is small.

The present invention is a kind of based on the adaptive steganography method for being directed to MP3 audio, main including secret information insertion and secret Confidential information extracts two parts.

Secret information embedded part occurs in MP3 audio encoding process after quantization encoding and before Huffman encoding, The embedding operation of information, including following sub-step are realized during audio coding:

1) sign bit for extracting qualified QMDCT coefficient in the area great Zhi and the area little Zhi of MP3 audio, believes as secret Cease the carrier of insertion；

2) according to the absolute hearing threshold in the modification amplitude and psychoacoustic model of QMDCT coefficient, calculating each can The modification cost of embedded location；

3) according to the modification cost of each embeddable position, selection is encoded using STC and is distorted the smallest insertion path, is led to The sign bit for crossing modification QMDCT coefficient completes the insertion of secret information.

The process of secret information insertion is further illustrated below by step S1~S6.

(S1) MP3 encoded audio frame includes two particles, and each particle includes 576 QMDCT coefficients, the area great Zhi and small value The coefficient in area requires to carry out Huffman encoding, in an encoding process, the sign bit of each nonzero coefficient is indicated with " 0 " or " 1 ". Embedding operation needs to change the sign bit of QMDCT coefficient, and QMDCT coefficient biggish for absolute value changes its sign bit and causes Frequency coefficient change greatly, often cause more serious perceptual distortion.Therefore, in algorithm in such a way that threshold value T is set, The range for the QMDCT coefficient selected when control insertion.MP3 carrier audio is decoded, for each particle, is opened from the area great Zhi Begin to traverse QMDCT coefficient in sequence, extracts sign bit.If coefficient value is greater than threshold value, this coefficient symbols position is unavailable, skips This coefficient continues to traverse next coefficient.Finally the area great Zhi and the sign bit in the area little Zhi are spliced, obtain only comprising " 0 " and The binary bits string C of " 1 ", the carrier being embedded in using this vector as secret information.

(S2) according to the human auditory system threshold value in the modification amplitude and psychoacoustic model of QMDCT coefficient, each is calculated The insertion of embeddable position is distorted, and finally obtains the corresponding modification cost ρ of C.

(S3) binary bits string C and modification cost ρ are subjected to scramble using identical scrambling algorithm and scramble key.

(S4) insertion of secret information is carried out using information coding tool STC.According to repairing for calculated each frequency coefficient Change cost and secret information that present granule need to be embedded in, STC operation is done to selected insertion carrier, adaptively chooses binary system Several positions in Bit String C are modified to achieve the purpose that information is embedded in, and realize minimum insertion all frequency domain systems of carrier The sum of several modification cost obtains carrying close sign bit binary bits string S.

(S5) use inverse scrambling algorithm corresponding with step S3 and scramble key are random by S progress inverted, obtain S '.

(S6) after the completion of secret information insertion, the sign bit of audio QMDCT coefficient is redistributed according to S ', and continue MP3 coding, obtains carrying close audio.

The secret information extraction unit molecular steps are as follows:

(P1) MP3 is carried close audio to be decoded, extracts in the area each particle great Zhi and the area little Zhi absolute value and is less than etc. In the sign bit S of the non-zero QMDCT coefficient of T_i.If coefficient is ineligible, continue the sign bit for extracting next coefficient.Most After obtain carrying close QMDCT coefficient symbols position binary bits string S.

(P2) S ' is obtained to S progress scramble using with scrambling algorithm consistent during embedding operation and scramble key.

(P3) it obtains after carrying close sign bit binary bits string S ', extracts secret information binary bits string using STC M。

Accordingly with above method, the present invention also provides a kind of MP3 audio-frequency information based on sign bit adaptive feed-forward network is hidden Hiding system comprising:

Sign bit extraction module is responsible for qualified QMDCT coefficient in the area great Zhi and the area little Zhi that extract MP3 audio Sign bit, the carrier as secret information insertion；

Cost computing module is modified, the human ear being responsible in the modification amplitude and psychoacoustic model according to QMDCT coefficient is listened Feel threshold value, calculates the modification cost of each embeddable position；

Secret information is embedded in module, is responsible for the modification cost according to each embeddable position, encodes selection using STC and loses Very the smallest insertion path, the sign bit by modifying QMDCT coefficient complete the insertion of secret information.

Further, above system further includes secret information extraction module, is responsible for carrying out using methods described above secret Confidential information extracts.

The beneficial effect of the present invention compared with prior art is:

(1) concealment is higher.Steganographic algorithm in the present invention has adaptivity, by calculating repairing for each insertion point Change cost, encoded in conjunction with STC, selection is embedded in secret information in distortion cost the smallest insertion point.Therefore, with higher anti- Detection property.

(2) steganographic capacity is big.Traditional steganography method, as the maximum steganographic capacity of MP3Stego only has 154pbs or so (two-channel, 44.1KHz).Method in the present invention, 320kbps code rate audio maximum steganographic capacity can reach 12kbps, Maximum steganographic capacity can also reach 10kbps when 128kbps code rate.

Detailed description of the invention

Fig. 1 is secret information embedding grammar flow chart；

Fig. 2 is the area great Zhi and the area the little Zhi codeword structure figure of MP3 quantization parameter；The wherein area (a) Tu Wei great Zhi, (b) figure be The area little Zhi；Huffcode indicates Huffman code word, and sign_x indicates that the sign bit of first coefficient, linbits_x indicate first The position linbits of a coefficient, sign_y indicate that the sign bit of the second coefficient, linbits_y indicate the sign bit of second coefficient. Similar, sign_v, sign_w, sign_x, sign_y respectively indicate the sign bit of 4 area Ge little Zhi QMDCT coefficients.

Fig. 3 is secret information extraction method flow chart.

Explanation is embodied

In order to make the foregoing objectives, features and advantages of the present invention clearer and more comprehensible, below by specific embodiment and Attached drawing is described in further details the present invention.

The flow chart of embedded mobile GIS of the invention is as shown in Figure 1.The realization of embedded mobile GIS mainly includes following four module: The extraction of sign bit, distortion function calculate, STC is encoded and the modification of sign bit.Firstly, mentioning take large values area and the area little Zhi in order The sign bit of qualified QMDCT coefficient is as carrier；Secondly, being utilized in conjunction with cost function and secret information bit vectors STC selects the smallest insertion path of distortion, and the sign bit by modifying QMDCT coefficient completes secret information insertion.Insertion Algorithm can be modeled as:

STC (C, M, ρ)=S,

Qc×S→Qs

Wherein, C indicates that the original QMDCT coefficient symbols position bit vectors extracted, M indicate that secret information, ρ indicate insertion Distortion cost, S indicate to carry close QMDCT coefficient symbols position bit vectors, Q_cIndicate original coefficient vector, Q_sIndicate to carry close coefficient to Amount.

The specific implementation process of information embedding method:

Step 1, MP3 encoded audio frame include two particles, and each particle includes 576 frequency coefficients.In cataloged procedure Coefficient is divided into the area great Zhi, the area little Zhi and zero area in sequence.The present invention selects the QMDCT coefficient in the area great Zhi and the area little Zhi Sign bit as embedded domain, the codeword structure in the area great Zhi and the area little Zhi is as shown in Figure 2.MP3 carrier audio is decoded, is mentioned Take out the non-zero QMDCT coefficient x that absolute value in the area great Zhi and the area little Zhi of each particle is less than or equal to T_jSign bit C_i.If x_jGreater than zero and absolute value is less than T, then symbol position C_iEqual to 0.If x_jLess than zero and absolute value is less than T, then symbol position C_iIt is equal to 1.It thus is avoided that QMDCT coefficient modifying biggish for absolute value, causes more serious perceptual distortion.If coefficient does not meet embedding Enter condition then to skip, continues to extract next coefficient.Finally obtain original QMDCT coefficient symbols position bit vectors C.Public affairs can be used Formula indicates are as follows:

Step 2 is mainly influenced by two factors because of MP3 audio distortion: the modification amplitude and psychologic acoustics of QMDCT coefficient Human auditory system threshold value in model.Absolute hearing threshold describes in squelch, and a pure tone is heard by human ear to be needed to have Standby energy.It therefore, can be with absolute hearing threshold curve reaction human ear to each QMDCT coefficient x_jSusceptibility.J pairs of frequency The absolute hearing threshold t answered_jIt can be calculated with following formula, t_jSmaller expression human ear is more sensitive to the audio signal of this frequency range.When Modification occurs in the frequency range of human ear sensitivity, it is easier to cause to be distorted.

If necessary to modify QMDCT coefficient x_jSign bit, x_jValue will become the opposite number-x being worth originally_j, use d_jTable The modification amplitude for showing coefficient value, then have

d_j=2 × | x_j|

Finally each embeddable position x according to cost function calculation_jInsertion be distorted ρ_j, finally obtain the corresponding distortion of C The function of ρ can be with is defined as:

Wherein, σ is constant to ensure (t_j+ σ) it is greater than 1；Logarithm operation plays smoothing effect, reduces the shadow of some extremums It rings.

Step 3: C and ρ is carried out scramble using identical scrambling algorithm and scramble key, C ' and ρ ' are obtained.In the present invention Identical scramble is carried out to C and ρ using Chaotic Scrambling method, the cost of insertion point is uniformly distributed.

Information insertion is carried out Step 4: encoding using STC.STC coding can be according to embeddable position x_jModification cost, Partial secret information is embedded into sign bit bit vectors C by adaptive selection optimal embedding path, and makes insertion total Distortion is minimum, obtains carrying close sign bit bit vectors S.Telescopiny can be formulated:

S=STC (C, M, ρ, height, width)

Wherein, height and width is the height and width of submatrix in STC cataloged procedure respectively.

Step 5: S is carried out inverted unrest using inverse scrambling algorithm corresponding with step 3 and scramble key, obtain S′。

Step 6: by the close sign bit S ' of load in S '_iIt is reassigned to non-zero and absolute value is less than or equal to threshold value T's QMDCT coefficient x_j.If S_iEqual to 1, then x is enabled_jSign bit be negative.If S_iEqual to 0, then x is enabled_jSign bit be positive.Finally MP3 coding is re-started, obtains carrying close audio.It is formulated are as follows:

As shown in figure 3, the secret information extraction unit molecular steps are as follows:

MP3 is carried close audio and is decoded by step 1, is extracted absolute value in the area each particle great Zhi and the area little Zhi and is less than The sign bit S of non-zero QMDCT coefficient equal to T_i.If coefficient is ineligible, continue the sign bit for extracting next coefficient. It finally obtains and carries close QMDCT coefficient symbols position bit vectors S.

Step 2: being obtained using with scrambling algorithm consistent during embedding operation and scramble key to S progress scramble S′。

Step 3: obtaining after carrying close sign bit bit vectors S ', secret information bit vectors M is extracted using STC decoding, The parameter that extraction process the uses such as width of generator matrix and height are consistent with parameter used in telescopiny.

M=STC (S, height, width)

Explanation will be further analyzed to technical advantage of the invention by specific experimental data below:

1, experimental setup

A length of 10s, two-channel MP3 audio each 1000 that code rate is 128kbps and 320kbps, make when choosing in experiment MP3 codec is Lame-3.99.5.Submatrix height is 7 in telescopiny, and width is respectively set to 2,4,10；System Number threshold value T is 2 and 4, and σ is set as 10 in cost function.

2, control experiment is arranged

Control experiment has selected other 3 kinds typical MP3 audio steganography algorithms, respectively Yan Diqun et al. propose based on The rear-mounted adaptive steganography in the area little Zhi that steganographic algorithm (Human Code Mapping, HCM), the Yao of code word replacement propose is calculated Method (Adaptive Post-Steganography, APS) and Yang Kun et al. propose based on isometric entropy coding replacement from It adapts to steganographic algorithm (Equal Length Entropy Codes Substitution, EECS).

3, steganalysis algorithm and the not setting of sentience measurement

Experiment uses the audio hidden information analysis algorithm (bibliography: Jin C, Wang R D, Yan based on Markov feature D Q.Steganalysis of MP3Stego with low embedding-rate using Markov feature[J] .Multimedia Tools&Applications, 2016,76 (5): 1-16.) and based on JRM feature detection algorithm (reference Document: J,Fridrich J.Steganalysis of JPEG images using rich models[J] .Media Watermarking, Security, and Forensics, 2012,8303:7-20.) carry out safety evaluation.It is real It tests middle use ODG value (Objective Difference Grade) and not sentience analysis is carried out to steganographic algorithm.

4, embedding capacity is analyzed

Experimental result in table 1 is all under conditions of insertion rate is equal to 0.5, and embedding capacity unit is bits per second (bps).As it can be seen from table 1 all there is biggish insertion to hold for HCM algorithm, EECS algorithm and ACS algorithm proposed by the present invention Amount, and this 3 kinds of algorithm capacity are relatively.But the embedding capacity of the ACS algorithm when threshold value T is equal to 4 is maximum, and The embedded quantity of ACS algorithm is determined by threshold value T, and the more big then steganographic capacity of threshold value is bigger.

The steganographic capacity of 1. algorithm of table compares

	HCM	APS	EECS	ACS (T=2)	ACS (T=4)
						128kbps	6069	573	5814	7647	10141
320kbps	10540	644	11008	9685	13142

5, sentience is not analyzed

The value range of ODG value between -4 to 0, ODG value is bigger indicate audio to be evaluated with referring to audio similarity more Height, the not sentience being distorted after steganography modification are better.Particularly, when two audio similarities are very high, ODG value can be big In 0.It is the WAV audio decompressed by MP3 bearer documents referring to audio in experiment, audio to be measured is to carry ciphertext part solution by MP3 Press obtained WAV audio.

Table 2 and table 3 are the ODG value in 128kbps and 320kbps subaudio frequency respectively.Wherein, " -- " indicate embedded mobile GIS without Method realizes this embedded quantity.It can be seen from the results that the ODG value of the ACS algorithm of the invention in 128kbps is all higher than other 3 kinds Algorithm.In 320kbps, ACS algorithm and EECS algorithm ODG value are closer to.Therefore, algorithm proposed by the present invention can not perceive Property is best.

ODG value compares under the conditions of 2. 128kbps of table

ODG value compares under the conditions of 3. 320kbps of table

6, anti-inspection property analysis

Negative sample and positive sample each 200 are used in experiment, wherein 70% is used as training sample, 30% is used as test Sample, classifier use SVM classifier.By the experimental data in table 4, it is apparent that the present invention proposes in 128kbps The accuracy that is detected of ACS method and EECS algorithm it is close.And in 320kbps, what ACS algorithm of the invention was detected Accuracy is minimum, and verification and measurement ratio is below 70% in the case of two kinds.It is possible thereby to illustrate, it is higher that the present invention proposes that algorithm has Safety.

Table 4. is compared using detection accuracy when Markov feature and JRM feature

The above embodiments are merely illustrative of the technical solutions of the present invention rather than is limited, the ordinary skill of this field Personnel can be with modification or equivalent replacement of the technical solution of the present invention are made, without departing from the spirit and scope of the present invention, this The protection scope of invention should be subject to described in claims.

Claims (10)

A method for hiding an MP3 audio information based on symbol bit adaptive embedding, comprising the steps of: 1) extracting the sign bit of the qualified QMDCT coefficient in the large value area and the small value area of the MP3 audio as a carrier embedded in the secret information; 2) Calculating the modification cost of each embeddable position according to the modification range of the QMDCT coefficient and the human auditory threshold in the psychoacoustic model; 3) According to the modification cost of each embeddable position, the embedded path with the least distortion is selected by STC coding, and the embedding of the secret information is completed by modifying the sign bit of the QMDCT coefficient. The method according to claim 1, wherein step 1) controls the range of QMDCT coefficients selected when embedding the secret information by setting a threshold to avoid serious modification of the QMDCT coefficients with large absolute values. Perceptual distortion. 3. The method of claim 2 wherein step 1) comprises: 1.1) decoding the MP3 carrier audio, extracting the sign bit C _i of the non-zero QMDCT coefficient x _j whose absolute value in the large value area and the small value area of each particle is less than or equal to the threshold T; 1.2) If x _{j is} greater than zero and the absolute value is less than T, the sign bit C _i is equal to 0; if x _{j is} less than zero and the absolute value is less than T, the sign bit C _i is equal to 1; if the coefficient does not meet the embedding condition, skip, Continue to traverse the next coefficient; 1.3) Splicing the sign bit of the large value area and the small value area to obtain a binary bit string C containing only 0 and 1 as a carrier embedded in secret information. 4. The method of claim 1 wherein step 2) uses the following formula to calculate the modification cost of the embeddable location: Where x _j represents the QMDCT coefficient; d _j represents the modified amplitude of the coefficient value; t _j represents the absolute auditory threshold; σ is a constant to ensure that (t _j + σ) is greater than 1; the logarithmic operation smoothes. The method according to claim 1, wherein in step 3), the STC encoding adaptively selects an optimal embedded path to embed part of the secret information into the symbol bit vector according to the modification cost of the embeddable position, and The total distortion of the embedding is minimized, and a symbolic bit vector of the payload is obtained. The method according to claim 1, wherein the step 3) sets the sign bit of the QMDCT coefficient extracted in step 1) and the modification cost calculated in step 2) using the same scrambling algorithm and scrambling key. Chaos, so that the cost of the embedded points can be evenly distributed, and then the secret information is embedded using STC coding, and then reversed using the corresponding inverse scrambling algorithm and scrambling key. The method according to claim 6, wherein after the secret information is embedded, the sign bit of the audio QMDCT coefficient is re-allocated according to the result obtained by the inverse scrambling, and the MP3 encoding is continued to obtain the encrypted audio. 8. Method according to claim 6, characterized in that the secret information is extracted using the following steps: (1) Decoding the MP3 carrier audio, extracting the sign bit of the non-zero QMDCT coefficient whose absolute value in the large value area and the small value area of each particle is less than or equal to the threshold value; if the coefficient does not meet the condition, continue to extract the next coefficient Sign bit; finally get the dense QMDCT coefficient sign bit binary bit string S; (2) scrambling S by using a scrambling algorithm and a scrambling key consistent with the embedding operation to obtain a binary bit string S' of the payload symbol bit; (3) The secret information binary bit string is extracted by STC decoding according to the hard coded bit binary bit string S'. 9. An MP3 audio information hiding system based on symbol bit adaptive embedding, comprising: The symbol bit extraction module is responsible for extracting the sign bit of the qualified QMDCT coefficient in the large value area and the small value area of the MP3 audio, as a carrier embedded in the secret information; Modifying the cost calculation module, responsible for calculating the modification cost of each embeddable position according to the modification range of the QMDCT coefficient and the human auditory threshold in the psychoacoustic model; The secret information embedding module is responsible for selecting the embedding path with the least distortion according to the modification cost of each embeddable position, and embedding the secret information by modifying the sign bit of the QMDCT coefficient. 10. The system of claim 9, further comprising a secret information extraction module responsible for extracting secret information using the following steps: (1) Decoding the MP3 carrier audio, extracting the sign bit of the non-zero QMDCT coefficient whose absolute value in the large value area and the small value area of each particle is less than or equal to the threshold value; if the coefficient does not meet the condition, continue to extract the next coefficient Sign bit; finally get the dense QMDCT coefficient sign bit binary bit string S; (2) scrambling S by using a scrambling algorithm and a scrambling key consistent with the embedding operation to obtain a binary bit string S' of the payload symbol bit; (3) The secret information binary bit string is extracted by STC decoding according to the hard coded bit binary bit string S'.