Robust audio steganography method and system for resisting social network channel transcoding

Abstract

The invention discloses a robust audio steganography method and system for resisting social network channel transcoding, and belongs to the technical field of information security. The method comprises the following steps: generating a carrier bit sequence based on a QMDCT coefficient set of an original audio; obtaining a cost sequence based on the modification cost of each bit in the carrier bit sequence; generating an error correction coding result of the secret information to be embedded; synchronously scrambling the carrier bit sequence and the cost sequence to obtain an out-of-order carrier bit sequence and an out-of-order cost sequence; searching a path with the minimum total modification cost in a grid chart defined by the out-of-order carrier bit sequence and the error correction coding result based on the out-of-order cost sequence to obtain an out-of-order secret-carrying bit sequence; and generating a secret-carrying audio based on the out-of-order secret-carrying bit sequence. According to the method, on the premise of not sacrificing excessive embedding capacity, an extremely low bit error rate and excellent imperceptibility and steganalysis resistance can still be kept in a lossy channel.

Description

Technical Field

[0001] This invention relates to the field of information security technology, and in particular to a robust audio steganography method and system that resists social network channel transcoding. Background Technology

[0002] The application and popularity of internet social media platforms (such as WeChat, Weibo, Bilibili, and Facebook) have provided new opportunities and transmission channels for the development and utilization of information hiding technologies. Currently, many active audio sharing platforms exist within social networks. These platforms all possess the advantages of a shared space model, providing protection for "one-to-many" communication relationships. However, to save network transmission bandwidth and storage space for streaming media, these platforms typically perform compression, transcoding, and format conversion on uploaded audio files. Furthermore, the encoder parameters and processing methods vary significantly across different platforms, and even the encoding processing modes used on the same platform exhibit variability. These challenges place new and higher demands on the robustness, reliability, and security of audio steganography technologies in practical applications on social network channels.

[0003] Audio, as one of the most common multimedia file contents on the internet, is widely used. The abundance of online audio carriers provides a favorable environment and conditions for information hiding. Among them, Advanced Audio Coding (ACC) compression audio format, due to its high compression ratio and high sound quality, has become one of the most popular audio coding formats and is a relatively ideal steganography embedding carrier. Current audio steganography research mainly focuses on the design of adaptive steganography methods under lossless channels, with very few research results on lossy channels. A novelty search of robust audio hiding techniques revealed that representative techniques both domestically and internationally include: Li et al. proposed a cepstral-based steganography method, which has certain robustness, but the embedding capacity is only 20-40 bps (Reference: Li X, Yu H. Transparent and robust audio datahiding in subband domain[J]. IEEE Proceedings International Conference on Information Technology: Coding and Computing, 2000: 74-79); For MDCT (Modified Discrete Cosine Transform) coefficients, a robust steganography algorithm based on MP3 encoding, called RMSM, was extracted, but it ignores the existence of silent audio segments (Reference: Yang Y, YiX, Zhao X, et alA robust MP3 steganographic method against multiple compressions based on modified discrete cosine transform[J]. Security and Communication Networks, 2021, 2021(1): 4471117; Another study proposed a robust steganography method for the Singular Value Decomposition (SVD) transform domain, which uses speech frame detection to filter embedded frames, but lacks consideration for steganographic perturbations (Reference: Kanhe A, Aghila GA DCT–SVD-based speech steganography in voiced frames[J]. Circuits, systems, and signal processing, 2018, 37(11): 5049-5068).

[0004] Research on robust audio hiding technologies reveals that representative invention patents currently available include: Chinese invention patent CN115620733A, which discloses "An AAC Secure Steganography Algorithm and System Based on Masking Effect," which analyzes the quantization process of MDCT coefficients in AAC encoding, selects the ear masking region as the embedding location, and combines it with an STC adaptive framework to achieve information hiding, improving security while ensuring auditory concealment; and Chinese invention patent CN117496986A, which discloses "Information Steganography Embedding and Extraction Method and System for AAC Bitstream Based on Semi-Decoding," which uses semi-decoding technology to parse the AAC bitstream down to the QMDCT coefficient level for embedding operations, significantly improving steganography efficiency and optimizing embedding capacity.

[0005] However, the design and implementation of steganography methods require comprehensive consideration of factors such as embedding capacity, undetectability, and imperceptibility. Despite the multi-round, intensive lossy processing employed by social media platforms, it remains difficult to guarantee the integrity of the secret information. This series of operations severely alters or even completely destroys the underlying data structure of the audio signal, causing significant distortion or complete loss of the embedded secret information. Analysis reveals the following limitations of existing robust audio steganography methods: (1) Existing time-domain steganography and compressed-domain steganography algorithms are generally designed based on lossless channels, which are difficult to resist the disturbances caused by the re-compression of social network channels.

[0006] (2) Although some existing transform domain steganography algorithms have a certain degree of robustness, they often lack consideration for perceptual distortion in their design. They have poor security.

[0007] (3) Existing robust audio steganography algorithms have small embedding capacity and are not robust enough. After being recompressed through social network channels, the bit error rate is high, and there is considerable room for improvement. Summary of the Invention

[0008] To address the problems of poor robustness, high bit error rate, and easy introduction of perceptual noise in existing audio steganography methods under lossy channels such as social networks, this invention proposes a robust audio steganography method and system that resists transcoding in social network channels. It achieves an extremely low bit error rate in lossy channels of mainstream social network platforms without sacrificing too much embedding capacity, while ensuring excellent imperceptibility and resistance to steganalysis.

[0009] To achieve the above objectives, the technical solution of the present invention includes the following:

[0010] An audio steganography method resistant to social network channel transcoding, the method comprising: Based on the QMDCT coefficient set of the original audio, a carrier bit sequence is generated; Based on the modification cost of each bit in the carrier bit sequence, the cost sequence is obtained; Generate the error-correcting encoding result of the secret information to be embedded; Synchronize and scramble the carrier bit sequence and the cost sequence to obtain the disordered carrier bit sequence and the disordered cost sequence; Based on the out-of-order cost sequence, the path with the minimum total modification cost is searched in the grid graph defined by the out-of-order carrier bit sequence and the error correction coding result to obtain the out-of-order carrier bit sequence. Based on the disordered encrypted bit sequence, encrypted audio is generated.

[0011] Furthermore, the process of generating the QMDCT coefficient set includes: The original video is segmented to obtain several audio frames; Calculate the perceptual entropy for each audio frame; Perform MDCT transformation on each audio frame to obtain the MDCT coefficients of that audio frame; Based on the perceptual entropy and the set audio bitrate and sampling rate, the MDCT coefficients are quantized to obtain the QMDCT coefficient set of the original video.

[0012] Furthermore, based on the QMDCT coefficient set of the original audio, a carrier bit sequence is generated, including: The QMDCT coefficient set is divided into multiple sub-bands; The quantized MDCT coefficients within each subband are divided into an even-numbered coefficient subset and an odd-numbered coefficient subset; The binary bits of this subband are obtained by comparing the maximum values ​​of the repeated coefficients in the even-numbered coefficient subset and the odd-numbered coefficient subset. By concatenating the binary bits of all sub-bands, the carrier bit sequence is obtained.

[0013] Furthermore, based on the modification cost of each bit in the carrier bit sequence, a cost sequence is obtained, including: For a subband where all quantized MDCT coefficients are zero, set the modification cost of the corresponding bits of that subband to infinity; For at least one subband with non-zero quantized MDCT coefficients, the modification cost of the bits corresponding to the remaining subbands is obtained by calculating the perturbation cost and the robustness cost; wherein the perturbation cost is proportional to the magnitude of the modification coefficient, and the robustness cost is obtained based on the maximum magnitude value of all quantized MDCT coefficients in the subband.

[0014] Furthermore, based on the out-of-order encrypted bit sequence, encrypted audio is generated, including: Restore the disordered bit sequence to an ordered bit sequence; The bits that need to be modified in the ordered bit sequence are determined, and the quantization MDCT coefficients in the subband corresponding to the bit that needs to be modified are adjusted to generate a set of QMDCT coefficients. The encrypted audio is obtained by inverse quantization and inverse MDCT transform of the encrypted QMDCT coefficient set.

[0015] A method for extracting secret information that resists social network channel transcoding, the method comprising: Obtain encrypted audio using any of the methods described above; Perform LMDCT transform on the encrypted audio to obtain the MDCT coefficient set; An ordered dense bit sequence was reconstructed based on the MDCT coefficient set; Scramble the ordered dense bit sequence to obtain the disordered dense bit sequence. Decode the out-of-order secret bit sequence to obtain the original secret information.

[0016] Furthermore, an ordered dense bit sequence is reconstructed based on the MDCT coefficient set, including: Merge the MDCT coefficients of all audio frames in sequence. The merged result is divided into several sub-bands, and the MDCT coefficients in each sub-band are divided into an even-numbered coefficient subset and an odd-numbered coefficient subset; The binary bits of this subband are obtained by comparing the maximum values ​​of the repeated coefficients in the even-numbered coefficient subset and the odd-numbered coefficient subset. By concatenating the binary bits of all sub-bands, the carrier bit sequence is obtained.

[0017] A robust audio steganography system resistant to social network channel transcoding, the system comprising: The carrier bit sequence generation module is used to generate a carrier bit sequence based on the QMDCT coefficient set of the original audio. The cost sequence generation module is used to obtain the cost sequence based on the modification cost of each bit in the carrier bit sequence; The encoding module is used to generate error-correcting encoding results for the secret information to be embedded. The first sequence scrambling module is used to synchronize and scramble the carrier bit sequence and the cost sequence to obtain the disordered carrier bit sequence and the disordered cost sequence. The secret-carrying sequence generation module is used to search for the path with the minimum total modification cost in the grid graph defined by the disordered carrier bit sequence and the error correction coding result based on the disordered cost sequence, so as to obtain the disordered carrier bit sequence. The encrypted audio generation module is used to generate encrypted audio based on the out-of-order encrypted bit sequence.

[0018] A secret information extraction system resistant to social network channel transcoding, characterized in that the system comprises: A encrypted audio acquisition module is used to acquire encrypted audio obtained based on any of the methods described above; The LMDCT transform module is used to perform LMDCT transform on the encrypted audio to obtain the MDCT coefficient set. The information reconstruction module is used to reconstruct an ordered dense bit sequence based on the MDCT coefficient set; The second sequence scrambling module is used to scramble the ordered dense bit sequence to obtain the disordered dense bit sequence. The decoding module is used to decode the out-of-order secret bit sequence to obtain the original secret information.

[0019] A computer device, characterized in that the computer device comprises: a processor and a memory storing computer program instructions; the processor, when executing the computer program instructions, implements the audio steganography method resisting social network channel transcoding or the secret information extraction method resisting social network channel transcoding as described above.

[0020] Compared with the prior art, the present invention has at least the following beneficial effects.

[0021] (1) In this invention, the steganography method has extremely high robustness. By fixing key parameters (window type) at the encoding level, the problem of inconsistent extraction environment caused by channel transcoding is fundamentally solved. Combined with a bit mapping strategy that guarantees the minimum amplitude difference, this method can effectively resist MP3 / AAC recompression and transcoding attacks from mainstream social platforms such as Bilibili and SoundCloud, achieving secret information transmission with zero bit error rate.

[0022] (2) Compared with existing robust steganography algorithms, the present invention has better imperceptibility. The joint cost function designed in this invention balances acoustic perturbation and robustness when calculating modification costs, and finds the optimal embedding path through STC encoding, thereby maintaining the excellent listening experience of the encrypted audio while ensuring strong robustness.

[0023] (3) Compared with the prior art, the present invention has higher security. Since the embedding position is adaptively selected based on the audio content, it disrupts the fixed statistical regularity. The recompression process itself also interferes with the original statistical feature distribution, making it difficult for steganalysis methods for specific compression formats to effectively capture steganalysis traces, thus possessing good anti-steganography analysis capabilities. Attached Figure Description

[0024] Figure 1 This is a flowchart of a robust audio steganography method that resists social network channel transcoding. Detailed Implementation

[0025] To make the various technical features, advantages, or effects of the present invention more apparent and understandable, a detailed description is provided below in conjunction with the accompanying drawings.

[0026] To resist lossy processing such as transcoding and recompression commonly found in social network channels and ensure reliable extraction of secret information, this invention deeply utilizes the inherent characteristics of ACC (Accelerated Cosine Transform). It fixes the long window type of MDCT (Modified Discrete Cosine Transform) and innovatively constructs a robust embedding domain and bit mapping rule based on the parity of the "maximum amplitude position" within the QMDCT (Quantized Modified Discrete Cosine Transform) coefficient block. This rule is insensitive to small-range fluctuations in coefficient values ​​caused by recompression. Furthermore, this invention designs a joint cost function that balances embedding distortion and channel robustness. This function effectively avoids the risks associated with embedding information in silent frames and guides subsequent Syndrome-Trellis Code (STC code) to adaptively embed based on the audio content, minimizing the overall modification cost. Finally, by combining BCH (Bose-Chaudhuri-Hocquenghem) error correction coding, this invention achieves an extremely low bit error rate in the lossy channels of mainstream social networking platforms without sacrificing too much embedding capacity, while ensuring excellent imperceptibility and resistance to steganalysis.

[0027] like Figure 1 As shown, the robust audio steganography method against social network channel transcoding of the present invention includes the following steps.

[0028] Step S1: Generate a carrier bit sequence based on the QMDCT coefficient set of the original audio.

[0029] Step S11: Frame segmentation of the bitstream. Original audio The data is segmented according to the AAC standard, and each frame typically has a data length of 2048 sampling points. This framing method is the basis for subsequent frequency domain transformation.

[0030] Step S12: Calculate the perceptual entropy of the audio frame. Calculate the perceptual entropy of the psychoacoustic model for each audio frame. This step follows the core principles of AAC encoding. For each audio frame, a built-in psychoacoustic model is used for analysis to calculate its masking threshold. Perceptual entropy is a key parameter derived from this masking threshold; it quantifies the degree to which the audio frame can be compressed. In the standard AAC encoding process, perceptual entropy... This will guide subsequent window selection, bit allocation, and quantization processes to achieve efficient compression while maintaining sound quality.

[0031] Further, in step S12, the perceptual entropy of the psychoacoustic model is calculated using the formula... We obtain, among which, perceptual entropy This will guide the quantization process. Function This table presents the psychoacoustic model defined in the AAC standard. This model simulates the human auditory masking effect by analyzing audio frames. Perform a short-time Fourier transform to analyze its spectral characteristics and calculate the masking threshold at different frequency points. (Perceptual entropy) It is a scalar value derived from these masking thresholds, quantifying the redundancy that an audio frame can be compressed, i.e., the amount of information that can be removed without introducing audible distortion. Therefore, perceptual entropy... It is a key input parameter in the subsequent adaptive quantization step S14.

[0032] Step S13: LMDCT Transform. To ensure consistent audio parsing at the receiving end, especially after lossy processing such as social network transcoding, this invention consistently uses a long window to perform MDCT (Improved Discrete Cosine Transform) on all audio frames. This results in high consistency of window type between the embedding and extraction ends, greatly enhancing the robustness of the steganography method. For each frame... After transformation, a set of MDCT coefficients containing 1024 frequency domain coefficients is obtained. .

[0033] Furthermore, in step S13, Through formula This leads to the adoption of a uniform long-block sinusoidal window sequence transform. Specifically, The transformation involves two core steps. First, it transforms audio frames with a length of 2048 samples... Apply a long block sine window function Its formula is: Signal after adding a window Secondly, regarding the windowed signal... Performing Discrete Cosine Transform (DCT) yields 1024 MDCT coefficients. The calculation formula is as follows: Where 0 ≤ k < 1024 This invention consistently employs this long window transform to ensure consistency between the embedding and extraction ends in the transform domain, thereby effectively resisting window type switching issues that may be caused by channel transcoding and significantly enhancing the robustness of the steganography method.

[0034] Step S14: Extract the quantization coefficient set. For MDCT coefficients... Quantization is performed. The quantization step size during the quantization process is determined by the perceptual entropy. The parameters, including the target bit rate, are jointly determined, with the aim of representing MDCT coefficients using as few bits as possible without introducing audible distortion. The quantized coefficients are then called QMDCT coefficients. The QMDCT coefficients obtained after processing all audio frames. The coefficients are merged sequentially to form a complete set. This set is the carrier field for all subsequent steganography operations.

[0035] Furthermore, in step S14, Through formula get. The function represents a complex iterative quantization and bit allocation process. This process uses perceptual entropy. To guide, combine the target bitrate of the audio. and sampling rate , is the MDCT coefficient Different number of bits are allocated to different frequency bands (sub-bands), and the optimal quantization step size is determined. The goal is to minimize audible distortion by masking quantization noise as much as possible by the audible masking threshold while meeting the bit rate limit. The output of this process is the quantized MDCT coefficients. .

[0036] Step S15: Bit Mapping. This invention employs a mapping method based on the parity of the maximum amplitude position within a QMDCT coefficient subband. Specifically, the coefficient set... The system is divided into multiple subbands, for example, with groups of four coefficients each. Within each subband, based on the parity of the coefficient indices, it is further divided into an even-numbered coefficient subset and an odd-numbered coefficient subset. By comparing the maximum amplitude of the coefficients in these two subsets, a binary bit is generated according to a preset rule (e.g., if the maximum amplitude of the odd-numbered subset is greater than or equal to that of the even-numbered subset, it is mapped to a bit "1", otherwise it is mapped to a bit "0"). The mapping results of all subbands are then concatenated sequentially to obtain the bit sequence corresponding to the original carrier. ; Furthermore, in step S15, The corresponding bit sequence uses the formula For example, for a subband containing 4 QMDCTs. It can be divided into odd-numbered subsets. and even subband set Assuming The position of the maximum amplitude is , The position of the maximum amplitude is (When there are multiple maximum values, take the element with the smallest position), where the corresponding bit sequence c is calculated by the following formula: Step S2: Based on the modification cost of each bit in the carrier bit sequence, obtain the cost sequence.

[0037] To achieve adaptive steganography, a bit sequence is required. Each element in the algorithm is assigned a modification cost. This invention designs a joint cost function that combines imperceptibility and robustness: 1) For silent sub-bands in the audio (i.e., sub-bands where all QMDCT coefficient values ​​are 0), their modification cost is set to infinity to forcibly avoid embedding information there. 2) For non-silent sub-bands, the cost is jointly determined by the "perturbation cost" (proportional to the magnitude of the modified coefficients) and the "robust cost" (related to the original maximum coefficient value within the sub-band, prioritizing sub-bands with smaller maximum values ​​to obtain stronger resistance to re-compression). The cost sequence calculated by this function... This will guide subsequent STC encoding to minimize overall distortion.

[0038] Furthermore, in step S2, utilizing the concept of wet paper coding, this paper sets the sub-band with all QMDCT coefficient values ​​of 0 (i.e., the silent sub-band) as the wet point, avoiding modification of it, thereby reducing the negative impact of silent frames on the bit error rate. For a sub-band containing... Subband of root frequency line It is possible to design an all-zero subband detection function. Its formula is: In actual embedding, at most one coefficient per subband will be modified. Therefore, the modification magnitude of this coefficient will be directly used as the perturbation cost for the QMDCT coefficients before modification of a single subband. and the modified QMDCT coefficients Its disturbance cost Set to: The weights used to adjust the cost of disturbance.

[0039] Given that the amplitude difference is fixed at 2 in the modification scheme, this indicates that the larger the maximum amplitude of the sub-band, the lower the relative amplitude of its modification; however, its bit error rate may also increase accordingly. In lossy channel transmission, error correction codes are essential, therefore, it is not necessary to pursue the lowest possible bit error rate. A strategy that prioritizes embedding in sub-bands with relatively low amplitudes within a certain bit error rate can be chosen. For a maximum amplitude value of... The robustness cost of the subband is set as follows: in, For the optimal embedding value, and These are used to adjust the magnitude of the increase on the left and right sides of the optimal embedding value. .

[0040] Therefore, for a single QMDCT subband ,like Its modification cost is set as a robustness cost. With the cost of disturbance The product of.

[0041] For vector sequence and the dense sequence QMDCT subband matrix The final cost of modification is: Step S3: Generate the error correction coding result of the secret information to be embedded.

[0042] Considering that encrypted audio may introduce bit errors during transmission over channels such as social networks due to transcoding and downsampling, this step uses forward error-correcting codes such as BCH codes to handle the original secret message to be signed in. Encode the data. The resulting bit sequence is... It contains redundant information, enabling the receiving end to detect and correct a certain number of errors, thereby ensuring the reliable recovery of secret messages.

[0043] Step S4: Synchronize and scramble the carrier bit sequence and the cost sequence to obtain the out-of-order carrier bit sequence and the out-of-order cost sequence.

[0044] To enhance the security of the steganography system and prevent attackers from detecting steganographic content by analyzing data statistical characteristics, this step uses a key. The controlled scrambling algorithm applies to the carrier bit sequence and cost sequence Perform synchronous scrambling. After scrambling, adjacent bits and costs are physically shuffled, increasing the difficulty of cracking.

[0045] Furthermore, in step S4, the scrambled out-of-order carrier bit sequence and disordered cost sequence Using formulas respectively and get. The function (scrambling function) is a function that uses a key. A deterministic pseudo-random scrambling algorithm for initialization. It generates a sequence... (or The input is then reordered based on an index permutation sequence of equal length. and Using the exact same key and algorithm, the integrity of each bit is guaranteed. The cost After the chaos, The cost It remains its original corresponding value.

[0046] Step S5: Based on the out-of-order cost sequence, search for the path with the minimum total modification cost in the grid graph defined by the out-of-order carrier bit sequence and the error correction coding result to obtain the out-of-order carrier bit sequence.

[0047] This step is the core of adaptive embedding using STC encoding. The STC encoder uses an out-of-order carrier bit sequence. Messages to be embedded and disordered cost sequence As input, it utilizes the cost sequence. As a guideline, among all possible modification paths, find one that can embed the message. And the total modification cost (i.e., the cost sequence) The path that minimizes the sum of costs for the corresponding modified bits. The modified result of this path is the out-of-order sequence of data bits. .

[0048] Furthermore, in step S5, the STC encoder will be invoked, and the formula will be used. Calculation . This represents the embedding function for the check trellis coding. This function utilizes a cost sequence. As a metric for pathfinding algorithms, in the context of carriers and secret messages In the defined mesh graph, find a path that minimizes the total cost. This path uniquely defines the path to the desired outcome. The modified scheme, the final output This means that a secret message was embedded. The out-of-order bit sequence with the least total distortion.

[0049] Step S6: Generate encrypted audio based on the out-of-order encrypted bit sequence.

[0050] Step S61: Inverse scramble the bit and cost sequence. Use the same key as described above. And the corresponding reverse scrambling algorithm, for the disordered data-carrying bit sequence Perform the operation to restore it to an ordered bit sequence consistent with the original audio frame order. .

[0051] Furthermore, in step S61, the formula is used... Enter the same key and To obtain the bit sequence to be embedded . The function is the exact inverse of the scrambling function described above. It uses the same key. Generate the inverse permutation sequence and sort the sequence. The elements are rearranged back to their original positions, resulting in an ordered sequence of data bits consistent with the original audio frame order. .

[0052] Step S62: Bit Inverse Mapping. This step maps the modifications to the binary field back to the QMDCT coefficient field. This is done by comparing the ordered bit sequence bit by bit. and carrier bit sequence First, identify the bits that need modification. For each bit, find its corresponding QMDCT subband and, according to the reverse process of the mapping rule described above, fine-tune one or more QMDCT coefficient values ​​within that subband (e.g., increase one coefficient value by a fixed small difference while decreasing another by a corresponding value) to change the parity of the position of the largest amplitude value within the subband, thereby achieving the embedding of the target bit. After performing this operation on all bits that need modification, the final set of QMDCT coefficients for carrier quantization is obtained. .

[0053] Furthermore, in step S62, preliminary experiments revealed that maintaining the QMDCT coefficient difference with an amplitude of 2 can maintain good robustness and imperceptibility when the target embedded bits are... The carrier message bits are The specific inverse mapping strategy is shown in the following equation: in, and Indicates the magnitude before embedding. and This represents the modified amplitude of the embedding. The core of this strategy is: when it is necessary to flip bits (i.e., ... Parity is altered by changing the position of one of the maximum values ​​by adding a fixed difference (e.g., +2), while possibly adjusting another value. This occurs when the bits do not need to be flipped. To further enhance robustness, the difference between the two maximum values ​​is checked and ensured to be at least 2. Throughout the modification process, only the amplitude values ​​of the QMDCTs coefficients are adjusted, while their coefficient signs remain unchanged.

[0054] Step S63: Inverse quantization and inverse MDCT transform. This step is the final step in reconstructing the dense audio. First, the dense quantization QMDCT coefficient set is processed. Each frame in Perform inverse quantization to obtain MDCT coefficients. Next, the MDCT coefficients for each frame... Perform an inverse MDCT transform to convert the frequency domain back to the time domain, obtaining the modified audio frame. Finally, concatenate all the modified audio frames to obtain the final encrypted audio. .

[0055] Further, in step S63, using the formula right Perform inverse quantization to obtain the modified MDCT coefficients. Using formula right Perform an inverse MDCT transform to convert it back from the frequency domain to the time domain, resulting in the modified audio frame. Finally, the obtained audio frames are merged to obtain audio containing the secret message. As an inverse quantization process, it multiplies the quantized index value by the corresponding quantization step size to recover the MDCT coefficient value. The inverse MDCT transform converts the frequency domain coefficients back to the time domain signal, ultimately generating a continuous PCM waveform.

[0056] In addition, the present invention also provides a secret information extraction method corresponding to the above-mentioned audio steganography method, which is used to recover secret information from a steganographic AAC audio file. This method is the inverse process of the embedding method, and the specific steps are as follows.

[0057] P1: Stream framing. The receiver first frames the received encrypted audio, which may have been transmitted and transcoded through social network channels. The data is decoded (usually in MP3, AAC, or uncompressed WAV format) to obtain a PCM data stream. Then, using the same parameters as the embedding end, the PCM stream is divided into audio frames with a length of 2048 sample points. Maintaining consistency in framing methods is crucial for ensuring synchronization in subsequent frequency domain analysis.

[0058] P2: LMDCT Transform. To ensure consistency between the extraction and embedding processes in the transform domain, and to combat window type mismatch issues caused by transcoding, this step consistently uses the same long window function as the embedding end, applying it to each audio frame. Perform LMDCT. After transformation, each audio frame corresponds to a set of MDCT coefficients containing 1024 frequency domain coefficients. .

[0059] Furthermore, in step P2, the formula is used. right MDCT coefficients are obtained by performing long-window MDCT transformation. Using the exact same transformation as the embedding end is a prerequisite for ensuring accurate extraction of embedded information, thus guaranteeing the robustness of the steganography system.

[0060] P3: Bit Mapping. This step is the core of information extraction and demonstrates the robustness of the method in this invention. Research has found that directly extracting bits from the unquantized MDCT coefficients achieves a lower bit error rate compared to quantizing first, because the quantization process itself may cause adjacent MDCT coefficient values ​​to be normalized to the same value, thereby destroying the embedded parity information. The specific extraction operation is as follows: MDCT coefficients of all frames... The data is merged sequentially and divided into subbands, for example, in groups of four, in the same manner as the embedding end. Within each subband, it is further divided into even-numbered coefficient subsets and odd-numbered coefficient subsets. A binary bit is "read" by comparing the maximum amplitude of the MDCT coefficients in these two subsets and applying the same mapping rules as the embedding end (e.g., if the maximum amplitude of the odd-numbered subset is greater than or equal to that of the even-numbered subset, it is considered a bit "1", otherwise it is "0"). The bits extracted from all subbands are then concatenated in their original order to reconstruct a complete and ordered sequence of MDCT bits. .

[0061] Furthermore, in step P3, using Unquantified MDCT coefficients Directly extract the embedded bit sequence .function This represents the exact same bit mapping rule as the embedding end. This is done to avoid information loss that may occur during quantization, thereby maximizing the accuracy and robustness of the extraction.

[0062] P4: Bit sequence inversion and scrambling. In order to correctly perform STC decoding, the extracted ordered data carrier bit sequence must be... It restores the key to its out-of-order state at the time of embedding. Therefore, this step uses the exact same key as the one used for embedding. And scrambling algorithms, for sequences Perform a scrambling operation to obtain an out-of-order sequence of data bits. .

[0063] Furthermore, in step P4, the formula is used. To obtain the bit sequence before scrambling, the formula contains... The function is exactly the same as the scrambling function used in the embedding end. It uses the same key as the embedding end. Generate a deterministic index permutation sequence, and apply this sequence to the extracted ordered data-carrying bit sequence. Perform a reordering. The purpose of this operation is to... Restoring the system to its out-of-order state before STC encoder processing is a necessary step for proper STC decoding.

[0064] P5: STC Decoding. Decodes the out-of-order data bits. The input is fed into the STC decoder. The STC decoder executes the decoding algorithm, thus obtaining... The secret message sequence encoded by BCH was recovered from the data. .

[0065] Furthermore, in step P5, the formula is used. The STC decoded bit sequence is obtained in this formula. This represents the STC decoding function. This function decodes the data using an out-of-order sequence of data bits. As input, the decoding algorithm is executed from The secret message sequence used during embedding is recovered from the data. The output of this process... It is a secret message encoded with BCH.

[0066] P6: BCH Decoding. Because the encrypted audio may have undergone lossy compression during transmission, the extracted... There are a few bit errors in the data; these errors will be passed on after STC decoding. In this step, the BCH decoding algorithm is used to... Decoding is then performed. The BCH decoder can detect and correct bit errors within a certain threshold, thereby effectively recovering data bits damaged by channel noise. After decoding, the output is the final, error-free original secret message. This completes the entire extraction process.

[0067] The technical advantages of this invention will be further analyzed and explained below using specific experimental data: Two local recompression channels and two online audio sharing channels will be used to test the robustness of the method. The online audio sharing channels are the Bilibili audio sharing platform and the SoundCloud platform. The Bilibili audio sharing platform channel outputs 192kbps stereo AAC audio files, while the SoundCloud platform channel outputs 128kbps stereo MP3 audio files. Both online audio sharing channels support uploading WAV and MP3 audio files. The AHCM algorithm (from the paper: Yi X, Yang K, Zhao X, et al. AHCM: Adaptive Huffman code mapping for audio steganography based on psychoacoustic model[J]. IEEE Transactions on Information Forensics and Security, 2019, 14(8): 2217-2231.) and the RMSM algorithm (from the paper: Yang Y, Yi X, Zhao X, et al. A robust MP3 steganographic method against multiple compressions based on modified discrete cosine transform[J]. Security and Communication Networks, 2021, 2021(1): 4471117.) were used as comparison objects. For each generated encrypted audio, its bit error rate before BCH decoding during extraction after passing through a lossy channel was calculated. and the bit error rate after decoding Finally, the average channel bit error rate was calculated. and effective transmission rate Its definition is: The test results are shown in Table 1. The average bit error rate of the proposed method after recompression on the local channel and the online channel is [not specified]. Both are around 1%, and after using error correction codes, they can reliably transmit messages without loss, showing significantly better robustness than RMSM and AHCM algorithms.

[0068] Table 1. Experimental results on robustness The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Those skilled in the art can modify or make equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention. The scope of protection of the present invention should be determined by the claims.

Claims

1. An audio steganography method resistant to social network channel transcoding, characterized in that, The method includes: Based on the QMDCT coefficient set of the original audio, a carrier bit sequence is generated; Based on the modification cost of each bit in the carrier bit sequence, the cost sequence is obtained; Generate the error-correcting encoding result of the secret information to be embedded; Synchronize and scramble the carrier bit sequence and the cost sequence to obtain the disordered carrier bit sequence and the disordered cost sequence; Based on the out-of-order cost sequence, the path with the minimum total modification cost is searched in the grid graph defined by the out-of-order carrier bit sequence and the error correction coding result to obtain the out-of-order carrier bit sequence. Based on the disordered encrypted bit sequence, encrypted audio is generated.

2. The method according to claim 1, characterized in that, The process of generating the QMDCT coefficient set includes: The original video is segmented to obtain several audio frames; Calculate the perceptual entropy for each audio frame; Perform MDCT transformation on each audio frame to obtain the MDCT coefficients of that audio frame; Based on the perceptual entropy and the set audio bitrate and sampling rate, the MDCT coefficients are quantized to obtain the QMDCT coefficient set of the original video.

3. The method according to claim 1, characterized in that, Based on the QMDCT coefficient set of the original audio, a carrier bit sequence is generated, including: The QMDCT coefficient set is divided into multiple sub-bands; The quantized MDCT coefficients within each subband are divided into an even-numbered coefficient subset and an odd-numbered coefficient subset; The binary bits of this subband are obtained by comparing the maximum values ​​of the repeated coefficients in the even-numbered coefficient subset and the odd-numbered coefficient subset. By concatenating the binary bits of all sub-bands, the carrier bit sequence is obtained.

4. The method according to claim 3, characterized in that, Based on the modification cost of each bit in the carrier bit sequence, the cost sequence is obtained, including: For a subband where all quantized MDCT coefficients are zero, set the modification cost of the corresponding bits of that subband to infinity; For at least one subband with non-zero quantized MDCT coefficients, the modification cost of the bits corresponding to the remaining subbands is obtained by calculating the perturbation cost and the robustness cost; wherein the perturbation cost is proportional to the magnitude of the modification coefficient, and the robustness cost is obtained based on the maximum magnitude value of all quantized MDCT coefficients in the subband.

5. The method according to claim 1, characterized in that, Based on the out-of-order encrypted bit sequence, encrypted audio is generated, including: Restore the disordered bit sequence to an ordered bit sequence; The bits that need to be modified in the ordered bit sequence are determined, and the quantization MDCT coefficients in the subband corresponding to the bit that needs to be modified are adjusted to generate a set of QMDCT coefficients. The encrypted audio is obtained by inverse quantization and inverse MDCT transform of the encrypted QMDCT coefficient set.

6. A method for extracting secret information that resists social network channel transcoding, characterized in that, The method includes: Obtain encrypted audio based on the method described in any one of claims 1 to 5; Perform LMDCT transform on the encrypted audio to obtain the MDCT coefficient set; An ordered dense bit sequence was reconstructed based on the MDCT coefficient set; Scramble the ordered dense bit sequence to obtain the disordered dense bit sequence. Decode the out-of-order secret bit sequence to obtain the original secret information.

7. The method according to claim 6, characterized in that, An ordered, dense bit sequence was reconstructed based on the MDCT coefficient set, including: Merge the MDCT coefficients of all audio frames in sequence. The merged result is divided into several sub-bands, and the MDCT coefficients in each sub-band are divided into an even-numbered coefficient subset and an odd-numbered coefficient subset; The binary bits of this subband are obtained by comparing the maximum values ​​of the repeated coefficients in the even-numbered coefficient subset and the odd-numbered coefficient subset. By concatenating the binary bits of all sub-bands, the carrier bit sequence is obtained.

8. A robust audio steganography system resistant to social network channel transcoding, characterized in that, The system includes: The carrier bit sequence generation module is used to generate a carrier bit sequence based on the QMDCT coefficient set of the original audio. The cost sequence generation module is used to obtain the cost sequence based on the modification cost of each bit in the carrier bit sequence; The encoding module is used to generate error-correcting encoding results for the secret information to be embedded. The first sequence scrambling module is used to synchronize and scramble the carrier bit sequence and the cost sequence to obtain the disordered carrier bit sequence and the disordered cost sequence. The secret-carrying sequence generation module is used to search for the path with the minimum total modification cost in the grid graph defined by the disordered carrier bit sequence and the error correction coding result based on the disordered cost sequence, so as to obtain the disordered carrier bit sequence. The encrypted audio generation module is used to generate encrypted audio based on the out-of-order encrypted bit sequence.

9. A secret information extraction system resistant to social network channel transcoding, characterized in that, The system includes: A encrypted audio acquisition module is used to acquire encrypted audio obtained by the method described in any one of claims 1 to 5; The LMDCT transform module is used to perform LMDCT transform on the encrypted audio to obtain the MDCT coefficient set. The information reconstruction module is used to reconstruct an ordered dense bit sequence based on the MDCT coefficient set; The second sequence scrambling module is used to scramble the ordered dense bit sequence to obtain the disordered dense bit sequence. The decoding module is used to decode the out-of-order secret bit sequence to obtain the original secret information.

10. A computer device, characterized in that, The computer device includes: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, it implements the audio steganography method resisting social network channel transcoding as described in any one of claims 1-5 or the secret information extraction method resisting social network channel transcoding as described in any one of claims 6-7.

Images