End-to-end audio steganography method and system

Abstract

The present invention discloses an end-to-end audio steganography method and system. The method comprises: using single-channel audio as a carrier, using short-time Fourier transform to convert the one-dimensional wave vector of the carrier audio and the secret audio into the frequency domain; inputting the secret audio converted into the frequency domain into a reversible network to realize the preprocessing of the audio information; using a reversible neural network to hide and extract the preprocessed audio information. The present invention adopts a reversible neural network, which has a bidirectional mapping capability from input to output. This mapping is bijective, which means that each input uniquely corresponds to an output, and each output also uniquely corresponds to an input. The encoding and decoding of the reversible network use the same parameters, which can retain more input details. End-to-end audio steganography does not require preprocessing of the input data, and the data can be directly input into the model to obtain the output result, thereby improving the efficiency of audio steganography.

Description

Technical Field

[0001] The present invention relates to the field of information security technology, and more specifically, to an end-to-end audio steganography method and system. Background Art

[0002] With the popularization and application of multimedia technology and the rapid development of social networks. It is becoming more and more popular for people to communicate with multimedia devices such as mobile phones and computers. People spread a large amount of video, audio, image, text and other information on social networks every day through social media. Some of this information may contain some personal privacy issues, such as home address, personal ID number, etc. There is a risk of information leakage when publicly transmitting information on social networks through multimedia devices. In particular, some personal privacy issues and some important information may be stolen and used by criminals when they are spread on the Internet, causing property losses. Communication security protection is mainly divided into encryption and information hiding: encryption mainly operates on the secret audio itself, but the plaintext after special processing is more susceptible to suspicion from third parties. Information hiding hides the existence of secret data, allowing secret data to be communicated covertly without causing suspicion from third parties. Therefore, information hiding, a communication security protection with camouflage characteristics, is increasingly used. Steganography is a technology that embeds secret audio in multimedia carriers without destroying the various properties of the carrier itself as much as possible. The most important feature of steganography is undetectability, and its purpose is to enable the communicating parties to communicate covertly without being detected by other users. Audio steganography is an important branch of steganography. Since a piece of audio can contain a lot of information, it is difficult to detect when hiding secret audio in it, making it an ideal secret audio carrier. Audio information transmitted through the network has become a new important carrier for implementing covert communication. Audio steganography that adapts to network channels is expected to become an important way to reliably and covertly transmit information in an open network environment.

[0003] At present, some audio steganography methods have been proposed. For example, the typical least significant bit replacement method combines discrete wavelet transform and fast Fourier transform to hide secret audio in speech. Recently, the rapid development of deep learning in the fields of computer vision and speech processing has achieved some relevant results in the field of digital image steganography. Among them, generative adversarial networks are used to generate carrier images with better embedding characteristics, and reversible neural networks are used to achieve end-to-end information steganography to achieve better steganography effects. The reversible process of the reversible block of the reversible neural network achieves a good secret audio hiding effect, and also has excellent results in resisting steganalysis attacks. Inspired by this, we try to combine reversible networks and adversarial training in audio steganography to achieve end-to-end audio steganography. Traditional machine learning methods often cannot directly use raw data, but need to process the raw data in advance, such as dimensionality reduction, feature extraction and other methods. The end-to-end audio steganography method we proposed can use the input data directly without other processing. Our model can automatically learn useful feature representations from raw data, reducing the need for human intervention and preprocessing. As far as we know, the research and application of end-to-end audio steganography is still in the pioneering stage.

[0004] Kuznetsov et al. studied the method of hiding information in audio using direct spread spectrum technology and explored the impact of five different ways of generating spread spectrum sequences on audio steganography (Kuznetsov, A., Onikiychuk, A., Peshkova, O., Gancarczyk, T., Warwas, K., & Ziubina, R. (2022). Direct spread spectrum technology for data hiding in audio. Sensors, 22 (9), 3115.)

[0005] Yang et al. proposed a framework based on generative adversarial networks to achieve optimal embedding in the time domain for audio steganography (Yang, J., Zheng, H., Kang, X., & Shi, YQ (2020, May). Approaching optimal embedding in audio steganography with GAN. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 2827-2831). IEEE.)

[0006] Nassrullah et al. proposed an efficient audio steganography method based on LSB, which improves the steganography performance by using the carrier to balance the hiding capacity and distortion rate (Nassrullah, HA, Flayyih, WN, & Nasrullah, MA (2020). Enhancement of LSB Audio Steganography Based on Carrier and Message Characteristics. J. Inf. Hiding Multim. Signal Process., 11(3), 126-137.)

[0007] Chen et al. proposed a technique for generating audio watermarks using a reversible network (Chen, G., Wu, Y., Liu, S., Liu, T., Du, X., & Wei, F. (2023). Wavmark: Watermarking for audio generation. arxiv preprint arxiv: 2308.12770.) Summary of the invention

[0008] In view of the above problems, an object of the present invention is to provide an end-to-end audio steganography method and system.

[0009] A first aspect of the present invention provides an end-to-end audio steganography method, the method comprising:

[0010] Using single-channel audio as the carrier, the one-dimensional wave vector of the carrier audio is converted into the frequency domain using short-time Fourier transform;

[0011] The one-dimensional wave vector of the secret audio is converted into the frequency domain using short-time Fourier transform;

[0012] The secret audio converted into the frequency domain and the carrier audio converted into the frequency domain are input into the reversible neural network, and the secret audio is embedded into the carrier audio by using the reversible neural network to obtain the steganographic audio;

[0013] Conduct simulated attacks on the stego-audio, then evaluate the difference between the carrier audio and the stego-audio and provide feedback on the performance of the reversible neural network,

[0014] Based on feedback about the performance of the reversible neural network, adjusting the reversible neural network to improve the accuracy of the audio generated by the generator to make it close to data from the original class;

[0015] The adjusted neural network is used to extract the steganographic audio after the simulated attack, and the steganographic secret audio and carrier audio are output.

[0016] Preferably, the formula for converting the one-dimensional wave vector of the carrier audio into the frequency domain using short-time Fourier transform is:

[0017] X specov =Γ stft (X cov )

[0018] Among them, X specov Frequency domain of carrier audio, X cov is the one-dimensional wave vector of the carrier audio, Γ stft () is the short-time Fourier transform formula.

[0019] Preferably, the formula for converting the one-dimensional wave vector of the secret audio into the frequency domain using short-time Fourier transform is:

[0020] X spesec =Γ stft (X sec )

[0021] Among them, X spesec is the frequency domain of the secret audio, X sec is the one-dimensional wave vector of the secret audio.

[0022] Preferably, the secret audio is embedded into the carrier audio, and the embedding process formula is:

[0023]

[0024] in, Indicates the hidden carrier audio. Indicates the hidden secret audio. It represents the Hadamard product operation, φ(.), ρ(.), and η(.) represent arbitrary functions. This application uses a 5-layer Denseblock to represent these three functions.

[0025] Preferably, the standardized formula for the operation taken in the process of reversely extracting the secret audio is:

[0026]

[0027] Among them, Z i Represents the extracted carrier audio, represents the extracted secret audio, Represents a matrix division operation.

[0028] The second aspect of the present invention provides an end-to-end audio steganography system, the system comprising: an STFT converter, a plurality of INN reversible modules, an ISTFT converter, a simulated attack module, and a discriminator;

[0029] The INN reversible module includes an information hiding reversible block and an information extraction reversible block;

[0030] The STFT transformer converts the one-dimensional wave vector of the carrier audio into the frequency domain, and the one-dimensional wave vector of the secret audio into the frequency domain;

[0031] The information hiding reversible block is used to generate a stego audio, embed the secret audio into the carrier audio, and obtain the loss and stego audio coefficients;

[0032] The information extraction reversible block is used to extract the secret information of the steganographic audio, and use the auxiliary information embedded in the embedding process to reverse the modification process and restore the original data of the secret information;

[0033] The ISTFT transformer is used to perform an inverse short-time Fourier transform on the steganographic audio coefficients to obtain the steganographic audio;

[0034] The simulated attack module is used to implement a simulated attack on the steganographic audio;

[0035] The discriminator is used to evaluate the difference between the carrier audio and the stego-audio and provide feedback on the performance of the information extraction reversible block, further the information extraction reversible block generates instances close to the secret data from the original class;

[0036] The STFT converter is also used to perform short-time Fourier transform on the attacked steganographic audio;

[0037] The information extraction reversible block reversely extracts the steganographic audio based on the steganographic audio information after short-time Fourier transformation combined with auxiliary information, and performs inverse short-time Fourier transformation on the extracted audio using the ISTFT transformer to obtain the steganographic carrier audio and secret audio.

[0038] Preferably, the information hiding reversible block modifies the original data in the secret audio during the process of embedding the secret audio into the carrier audio, and the modification can be reversible through the information extraction reversible block.

[0039] Preferably, the information extraction reversible block and the information hiding reversible block are reversible to each other.

[0040] Preferably, the discriminator consists of 6 groups connected in sequence, from group 1 to group 5, each group consists of a convolutional layer, a BN layer and a LeakyReLU activation function; the 6th group contains a global average pool and a linear layer for outputting classification probability.

[0041] Preferably, the convolutional layer of the discriminator has a convolution kernel size of 3, a stride of 2, and a padding of 1.

[0042] Compared with the prior art, the beneficial effect of the technical solution of the present invention is that the present invention provides an end-to-end audio steganography method and system. The present invention adopts a reversible neural network, which has a bidirectional mapping capability from input to output. This mapping is bijective, which means that each input uniquely corresponds to an output, and each output also uniquely corresponds to an input. The encoding and decoding of the reversible network use the same parameters, which can retain more input details. In addition, the present invention adds a discriminator module and an attack module to the overall architecture of the reversible neural network. The main function of the discriminator is to judge the direct similarity between the generated audio and the original carrier audio, adjust the generation effect of the generator according to the direct difference between the two, and continuously update the generator according to the principle of adversarial learning to generate audio that is closer to the original sound effect. The present invention also imitates the characteristics of ordinary attacks in the attack module, and the model adaptively learns the robustness to various attacks. End-to-end audio steganography does not require preprocessing of the input data, and the data can be directly input into the model to obtain the output result. Improve the efficiency of audio steganography. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a flow chart of an end-to-end audio steganography method described in Example 1.

[0044] Figure 2 This is a structural diagram of the reversible neural network described in Example 2.

[0045] Figure 3 This is a diagram of the reversible block structure of information hiding described in Example 2.

[0046] Figure 4 This is a reversible block structure diagram for information extraction described in Example 2.

[0047] Figure 5 This is a structural diagram of the identifier described in Example 2. DETAILED DESCRIPTION

[0048] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0049] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0050] Example 1

[0051] like Figure 1 As shown, this embodiment discloses an end-to-end audio steganography method. The method comprises:

[0052] S1: Using single-channel audio as the carrier, the one-dimensional wave vector of the carrier audio is converted into the frequency domain using short-time Fourier transform;

[0053] S2: Convert the one-dimensional wave vector of the secret audio into the frequency domain using short-time Fourier transform;

[0054] S3: input the secret audio converted into the frequency domain and the carrier audio converted into the frequency domain into the reversible neural network, and use the reversible neural network to embed the secret audio into the carrier audio to obtain the steganographic audio;

[0055] S4: Conduct simulated attacks on the stego-audio, then evaluate the difference between the carrier audio and the stego-audio and provide feedback on the performance of the reversible neural network,

[0056] S5: based on the feedback about the performance of the reversible neural network, adjusting the reversible neural network to improve the accuracy of the audio generated by the generator so that the audio generated by the generator is close to the data from the original class;

[0057] S6: Use the adjusted neural network to extract the steganographic audio after the simulated attack, and output the steganographic secret audio and carrier audio.

[0058] It should be noted that audio steganalysis is a technology that reveals secret information hidden by steganography, and audio steganography is a technology that hides secret information in digital carrier audio. In the past decade, many audio steganography and steganalysis methods have been proposed. With the development of artificial intelligence, audio steganalysis methods are constantly enhancing their detection capabilities. However, existing audio steganography methods mainly rely on manual features, which means that these methods are more easily detected to see if they are hiding information when fighting against steganalysis detection. The end-to-end audio steganography method proposed in this embodiment has better performance in terms of steganography effect and resistance to steganalysis detection. And compared with some existing audio steganography methods, the end-to-end audio steganography method proposed in the embodiment does not require a lot of preprocessing of the input data. The end-to-end audio steganography method can directly input data to obtain output.

[0059] In modern society, the importance of protecting digital data is very high. Using steganography is one of the methods to ensure the security of digital data. This is particularly relevant to various multimedia data, such as audio recordings, video files, etc. Using steganography to hide information is a mature and developing scientific field. With the development of social media and communication technology, people transmit a large amount of audio information through social networks every day, such as transmitting voice chat information through social software such as WeChat, QQ, Facebook, etc. When these voice messages are spread on social networks, important audio secret information can be hidden in the carrier through steganography methods to achieve the purpose of secretly spreading information. However, the current mainstream image steganography has a good effect in hiding images and text information. There are still many weaknesses in hiding a piece of audio. Audio steganography that hides audio with audio can achieve better results. Most of the current audio hiding technologies are based on traditional manual audio steganography methods, which perform poorly in terms of hiding effects, resistance to steganalysis detection, and the ability to suffer malicious attacks. The embodiment uses a reversible network to hide and extract audio, uses adversarial training to generate audio that is closer to the original sound, and uses a simulated attack module to improve the ability to resist ordinary attacks. Therefore, this embodiment uses end-to-end audio steganography with reversible networks and adversarial training to transmit secret information on social networks without being discovered.

[0060] According to an embodiment of the present invention, the formula for converting the one-dimensional wave vector of the carrier audio into the frequency domain using short-time Fourier transform is:

[0061] X specov =Γ stft (X cov )

[0062] Among them, X specov Frequency domain of carrier audio, X cov is the one-dimensional wave vector of the carrier audio, Γ stft () is the short-time Fourier transform formula.

[0063] According to an embodiment of the present invention, the formula for converting the one-dimensional wave vector of the secret audio into the frequency domain using short-time Fourier transform is:

[0064] X spesec =Γ stft (X sec )

[0065] Among them, X spesec is the frequency domain of the secret audio, S sec is the one-dimensional wave vector of the secret audio.

[0066] According to an embodiment of the present invention, the embedding process of the secret audio into the carrier audio is as follows:

[0067]

[0068]

[0069] in, Indicates the hidden carrier audio. Indicates the hidden secret audio. It represents the Hadamard product operation, φ(.), ρ(.), and η(.) represent arbitrary functions. This application uses a 5-layer Denseblock to represent these three functions.

[0070] According to an embodiment of the present invention, the standardized formula for the operation taken during the reverse extraction of secret audio is:

[0071]

[0072]

[0073] Among them, Z i Represents the extracted carrier audio, represents the extracted secret audio, Represents a matrix division operation.

[0074] In this embodiment, the secret audio and the carrier audio are subjected to the same short-time Fourier transform (STFT) as and Input into the hidden network, and use the hidden network to embed the secret audio into the carrier audio. According to the i-th reversible block in the reversible network, the embedding process can be reversibly converted into the following formula:

[0075]

[0076]

[0077] Here represents the Hadamard product operation. φ(.), ρ(.), η(.) represent arbitrary functions, which are not required to be reversible. In the embodiment, a 5-layer Denseblock is used in the entire network architecture to represent these three functions. The standardized formula for the operation taken in the process of reverse extraction of secret audio is as follows:

[0078]

[0079]

[0080] Among them, Z i Represents the extracted carrier audio, represents the extracted secret audio, Represents a matrix division operation.

[0081] Example 2

[0082] like Figure 2 As shown, this embodiment discloses an end-to-end audio steganography system, the system comprising: an STFT converter, a plurality of INN reversible modules, an ISTFT converter, a simulation attack module, and a discriminator;

[0083] The INN reversible module includes an information hiding reversible block and an information extraction reversible block;

[0084] The STFT transformer converts the one-dimensional wave vector of the carrier audio into the frequency domain, and the one-dimensional wave vector of the secret audio into the frequency domain;

[0085] The information hiding reversible block is used to generate a stego audio, embed the secret audio into the carrier audio, and obtain the loss and stego audio coefficients;

[0086] The information extraction reversible block is used to extract the secret information of the steganographic audio, and use the auxiliary information embedded in the embedding process to reverse the modification process and restore the original data of the secret information;

[0087] The ISTFT transformer is used to perform an inverse short-time Fourier transform on the steganographic audio coefficients to obtain the steganographic audio;

[0088] The simulated attack module is used to implement a simulated attack on the steganographic audio;

[0089] The discriminator is used to evaluate the difference between the carrier audio and the stego-audio and provide feedback on the performance of the information extraction reversible block, further the information extraction reversible block generates instances close to the secret data from the original class;

[0090] The STFT converter is also used to perform short-time Fourier transform on the attacked steganographic audio;

[0091] The information extraction reversible block reversely extracts the steganographic audio based on the steganographic audio information after short-time Fourier transformation combined with auxiliary information, and performs inverse short-time Fourier transformation on the extracted audio using the ISTFT transformer to obtain the steganographic carrier audio and secret audio.

[0092] According to an embodiment of the present invention, the information hiding reversible block modifies the original data in the secret audio during the process of embedding the secret audio into the carrier audio, and the modification can be reversible through the information extraction reversible block.

[0093] According to the embodiment of the present invention, the information extraction reversible block and the information hiding reversible block are reversible to each other.

[0094] It should be noted that INN is the abbreviation of Invertible Neural Networks, which is a reversible neural network. The INN reversible module is a reversible neural network module.

[0095] It should be noted that the traditional neural network uses two independent encoders and decoders to hide and extract information. The mapping relationship between the secret audio M, the carrier audio C, and the steganographic audio S can be expressed as: (M, C) → S and S → (M, C). This relatively independent structure of the encoder and decoder may accumulate the error of one mapping to another mapping. Based on this problem, this embodiment attempts to find a bijective network to complete the information hiding and extraction process with one network structure at the same time. The mapping relationship between reversible networks can be expressed as: Performing encoding and decoding operations in a network in this way can effectively avoid the occurrence of the above problems.

[0096] In this embodiment, 12 INN reversible modules are used in the reversible network for information hiding and extraction, and the INN reversible modules include information hiding reversible blocks and information extraction reversible blocks; parameters are shared during the training process of information hiding and extraction, while parameters of the network model of the traditional encoding-decoding architecture are not shared during the training process.

[0097] The attack module is a simulator that simulates the attacks that the steganographic audio information will suffer when it is transmitted on the network, and introduces it into the network to effectively improve the robustness of the system. Adversarial learning is used to distinguish whether the steganographic audio and the original audio are similar, and the steganographic audio effect is continuously improved through iterative training, ultimately achieving a good effect.

[0098] According to an embodiment of the present invention, the formula for converting the one-dimensional wave vector of the carrier audio into the frequency domain using short-time Fourier transform is:

[0099] X specov =Γ stft (X cov )

[0100] Among them, X specov Frequency domain of carrier audio, X cov is the one-dimensional wave vector of the carrier audio, Γ stft () is the short-time Fourier transform formula.

[0101] According to an embodiment of the present invention, the formula for converting the one-dimensional wave vector of the secret audio into the frequency domain using short-time Fourier transform is:

[0102] X spesec =Γ stft (X sec )

[0103] Among them, X spesec is the frequency domain of the secret audio, X sec is the one-dimensional wave vector of the secret audio.

[0104] It should be noted that, as a specific embodiment, this embodiment uses a single-channel audio with a sampling rate of 16kHz as a carrier, where the encoding length unit is set to one second. The original input of the carrier audio can be represented as a one-dimensional wave vector X with a length of 16000 cov Then, it is further converted into the frequency domain using short-time Fourier transform. The sampling rate is 16kHz for the secret information, where the encoding length unit is one second and the original input length is 16000. sec , and convert it into frequency domain using short-time Fourier transform. After this step, the secret information can be embedded into the carrier without causing a large transformation.

[0105] According to an embodiment of the present invention, the embedding process of the secret audio into the carrier audio is as follows:

[0106]

[0107]

[0108] in, Indicates the hidden carrier audio. Indicates the hidden secret audio. It represents the Hadamard product operation, φ(.), ρ(.), and η(.) represent arbitrary functions. This application uses a 5-layer Denseblock to represent these three functions.

[0109] According to an embodiment of the present invention, the standardized formula for the operation taken during the reverse extraction of secret audio is:

[0110]

[0111]

[0112] Among them, Z i Represents the extracted carrier audio, represents the extracted secret audio, Represents a matrix division operation.

[0113] It should be noted that the INN reversible module in this embodiment solves the embedding and extraction process of audio steganography in the same network. The INN reversible module is composed of several reversible blocks. The basic reversible coupling layer in the reversible network is the additive affine transformation. The INN reversible module includes an information hiding reversible block and an information extraction reversible block.

[0114] like Figure 3 The information hiding reversible block in the INN reversible module is shown. Figure 4 Extract reversible blocks for information in INN reversible modules. Figure 3 The i-th reversible block input is shown and The corresponding output is and Figure 4 Represents the reverse process.

[0115] In this embodiment, the secret audio and the carrier audio are subjected to the same short-time Fourier transform (STFT) as and Input into the hidden network, and use the hidden network to embed the secret audio into the carrier audio. According to the i-th reversible block in the reversible network, the embedding process can be reversibly converted into the following formula:

[0116]

[0117]

[0118] Here It represents the Hadamard product operation. φ(.), ρ(.), η(.) represent arbitrary functions, which are not required to be reversible. In this embodiment, the entire network architecture uses 5 layers of Denseblock to represent these three functions. The standardized formula for the operation taken in the process of reverse extraction of secret audio is as follows:

[0119]

[0120]

[0121] According to an embodiment of the present invention, the discriminator consists of 6 groups connected in sequence, from group 1 to group 5, each group consists of a convolutional layer, a BN layer and a LeakyReLU activation function; the 6th group contains a global average pool and a linear layer for outputting classification probability.

[0122] According to an embodiment of the present invention, the convolution kernel size of the convolution layer of the discriminator is 3, the stride is 2, and the padding is 1.

[0123] It should be noted that the discriminator is used to evaluate the difference between the carrier audio and the stego-audio and provide feedback on the performance of the generator, which further stimulates the generated instances to be closer to the data from the original class.

[0124] like Figure 5As shown, the discriminator in this embodiment consists of 6 groups. From group 1 to group 5, each group consists of a convolution layer (convolution kernel size = 3, stride = 2, padding = 1), a BN layer, and a LeakyReLU activation function. The 6th group contains a global average pooling (GAP) and a linear layer to output the classification probability. Using the basic principles of adversarial learning, the generated stego audio has high auditory fidelity and stego security against statistical detection.

[0125] As a specific embodiment, the reversible neural network of this embodiment uses small batch iterative stochastic gradient descent to train the cross entropy error. The adversarial training method is used to optimize the hyperparameters. Table 1 lists some important hyperparameters used to train the network. Under this structure, the proposed reversible network architecture model provides a relatively good steganographic effect.

[0126] Table 1 Training hyperparameters

[0127] Learning Rate <![CDATA[10 -6 ]]> Minimum batch size 8 <![CDATA[β1]]> 0.5 Training epoch 100 <![CDATA[β2]]> 0.999 Number of channels 1

[0128] Among them, β1 and β2 are ADAM optimizer parameters respectively.

[0129] In the specific experiment, the data set used for training and testing comes from TIMIT. In this embodiment, a total of 25,000 pairs of one-second audio are selected. During the training process, 20,000 pairs are used as carrier audio and secret information audio to train the entire network. 2,500 pairs are used for testing and 2,500 pairs are used for verification. The network is optimized using the Adam optimizer and standard parameters. The initial learning rate is set to 0.0001 and the batch size is set to 8. The entire training process includes 100 epochs. Two 4090 graphics cards with 24G video memory and three 3090 graphics cards with 24G video memory are used for the experiment. The environment of the corresponding deep learning framework is configured and correctly deployed on the server.

[0130] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0131] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units; they may be located in one place or distributed on multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0132] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0133] Those skilled in the art can understand that: all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiments; and the aforementioned storage medium includes: mobile storage devices, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), disks or optical disks, and other media that can store program codes.

[0134] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention can be essentially or partly reflected in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

Claims

1. An end-to-end audio steganography method, characterized in that: The method comprises: Using single-channel audio as the carrier, the one-dimensional wave vector of the carrier audio is converted into the frequency domain using short-time Fourier transform; The one-dimensional wave vector of the secret audio is converted into the frequency domain using short-time Fourier transform; The secret audio converted into the frequency domain and the carrier audio converted into the frequency domain are input into the reversible neural network, and the secret audio is embedded into the carrier audio by using the reversible neural network to obtain the steganographic audio; Conduct simulated attacks on the stego-audio, then evaluate the difference between the carrier audio and the stego-audio and provide feedback on the performance of the reversible neural network, Based on feedback about the performance of the reversible neural network, adjusting the reversible neural network to improve the accuracy of audio generated by the generator; Use the adjusted neural network to extract the steganographic audio after the simulated attack, and output the steganographic secret audio and carrier audio; The formula for converting the one-dimensional wave vector of the carrier audio into the frequency domain using short-time Fourier transform is: X specov =C stft (X cov ) Among them, X specov Frequency domain of carrier audio, X cov is the one-dimensional wave vector of the carrier audio, Γ stfr () is the short-time Fourier transform formula; The formula for converting the one-dimensional wave vector of secret audio into the frequency domain using short-time Fourier transform is: X spesec =C stft (X sec ) Among them, X spesec is the frequency domain of the secret audio, X sec is the one-dimensional wave vector of the secret audio; Embed the secret audio into the carrier audio. The embedding process formula is: Where, stft stands for short-time Fourier transform, i stands for the i-th reversible block, i belongs to 1-12, Indicates the hidden carrier audio. Indicates the carrier audio after stft. It represents the secret information audio after stft. Indicates the hidden secret audio. represents the Hadamard product operation, φ(.), ρ(.), η(.) represent arbitrary functions; The standardized formula for the operations taken during the reverse extraction of secret audio is: Where stft stands for short-time Fourier transform, i stands for the i-th reversible block, i belongs to 1-12, Z i Represents the extracted carrier audio, represents the extracted secret audio, Represents the matrix division operation, Z i+1 Represents auxiliary information that helps extract, Represents the extracted secret audio The previous process.

2. An end-to-end audio steganography system, characterized in that: The system includes: an STFT converter, a plurality of INN reversible modules, an ISTFT converter, a simulated attack module, and a discriminator; The INN reversible module includes an information hiding reversible block and an information extraction reversible block; The STFT transformer converts the one-dimensional wave vector of the carrier audio into the frequency domain, and the one-dimensional wave vector of the secret audio into the frequency domain; The information hiding reversible block is used to generate a stego audio, embed the secret audio into the carrier audio, and obtain the loss and stego audio coefficients; The information extraction reversible block is used to extract the secret information of the steganographic audio, and use the auxiliary information embedded in the embedding process to reverse the modification process and restore the original data of the secret information; The ISTFT transformer is used to perform an inverse short-time Fourier transform on the steganographic audio coefficients to obtain the steganographic audio; The simulated attack module is used to implement a simulated attack on the steganographic audio; The discriminator is used to evaluate the difference between the carrier audio and the stego-audio and provide feedback on the performance of the information extraction reversible block, further the information extraction reversible block generates instances close to the secret data from the original class; The STFT converter is also used to perform short-time Fourier transform on the attacked steganographic audio; The information extraction reversible block reversely extracts the steganographic audio according to the steganographic audio information after short-time Fourier transformation combined with the auxiliary information, and performs inverse short-time Fourier transformation on the extracted audio using the ISTFT transformer to obtain the carrier audio and secret audio after steganography; The formula for converting the one-dimensional wave vector of the carrier audio into the frequency domain is: X specov =C stft (X cov ) Among them, X specov Frequency domain of carrier audio, X cov is the one-dimensional wave vector of the carrier audio, Γ stfr () is the short-time Fourier transform formula; The formula for converting the one-dimensional wave vector of the secret audio into the frequency domain is: X spesec =C stft (X sec ) Among them, X spesec is the frequency domain of the secret audio, X sec is the one-dimensional wave vector of the secret audio; The secret audio is embedded in the carrier audio, and the embedding process formula is: Where, stft stands for short-time Fourier transform, i stands for the i-th reversible block, i belongs to 1-12, Indicates the hidden carrier audio. Indicates the carrier audio after stft. It represents the secret information audio after stft. Indicates the hidden secret audio. represents the Hadamard product operation, φ(.), ρ(.), η(.) represent arbitrary functions; The standardized formula for the operation adopted in the process of reverse extraction of the stego-audio is: Where stft stands for short-time Fourier transform, i stands for the i-th reversible block, i belongs to 1-12, Z i Represents the extracted carrier audio, represents the extracted secret audio, Represents the matrix division operation, Z i+1 Represents auxiliary information that helps extract, Represents the extracted secret audio The previous process.

3. The end-to-end audio steganography system according to claim 2, characterized in that: The information hiding reversible block modifies the original data in the secret audio during the process of embedding the secret audio into the carrier audio, and the modification can be reversible through the information extraction reversible block.

4. The end-to-end audio steganography system according to claim 3, characterized in that: The information extraction reversible block and the information hiding reversible block are reversible to each other.

5. The end-to-end audio steganography system according to claim 4, characterized in that: The discriminator consists of 6 groups connected in sequence, from group 1 to group 5, each group consists of a convolutional layer, a BN layer and a LeakyReLU activation function; the 6th group contains a global average pool and a linear layer for outputting the classification probability.

6. The end-to-end audio steganography system according to claim 5, characterized in that: The convolutional layer of the discriminator has kernel size = 3, stride = 2, and padding = 1.

Images