An improved swin transformer-based high-fidelity audio steganography method and system

Abstract

The application relates to a high-fidelity audio steganography method and system based on an improved Swin Transformer, and the method comprises the following steps: constructing an audio steganography network, wherein the audio steganography network comprises a hiding network and a recovery network; the hiding network and the recovery network both take an improved Swin Transformer as a core backbone; training the audio steganography network based on a training data set to obtain a trained hiding network and a trained recovery network; embedding secret information according to the trained hiding network and extracting secret information according to the trained recovery network. The application can realize high-capacity and high-recovery-quality secret audio information hiding while ensuring high host audio fidelity, and improves the concealment and reliability of steganography communication.

Description

Technical Field

[0001] This application relates to the field of information security, specifically to a high-fidelity audio steganography method and system based on an improved Swin Transformer. Background Technology

[0002] In the digital information age, the importance of information security is increasingly prominent. Steganography, as an important branch of information security, aims to embed secret information into public media (such as images, audio, and video) for transmission, thereby concealing the existence of the secret information and achieving covert communication. Unlike cryptography, which focuses on protecting the content of information, steganography aims to make the existence of secret information undetectable to attackers, thus circumventing censorship and attacks.

[0003] Existing deep learning-based steganography methods can learn the complex mapping relationship from the carrier to the secret information in an end-to-end manner, thereby achieving more efficient and covert information embedding. In particular, the introduction of generative adversarial networks (GANs) has greatly improved the perceptual quality of the steganographic carrier, making it more difficult to distinguish from the original carrier visually or aurally.

[0004] However, current research in the field of deep learning audio steganography mainly focuses on two directions: one is audio watermarking, which aims to embed a small amount of information to resist various attacks and is used for scenarios such as copyright protection; the other is cross-modal steganography, such as embedding image information into audio. In contrast, research on the specific task of "audio-to-audio" (that is, completely embedding a secret audio segment into another host audio segment), which has important application value (such as embedding commentary tracks in streaming media and adding supplementary explanations in educational audio), is still quite limited in the field of deep learning. The few existing exploratory works either rely on a simple combination of traditional LSB ideas and GANs, or their network structures are insufficient to effectively handle the complex time-frequency dependencies of audio spectrograms, making it difficult to achieve an ideal balance between embedding capacity, host fidelity, and the quality of secret information recovery. Summary of the Invention

[0005] To overcome at least one deficiency in the prior art, this application provides a high-fidelity audio steganography method and system based on an improved SwingTransformer.

[0006] Firstly, a high-fidelity audio steganography method based on an improved Swin Transformer is provided, including:

[0007] An audio steganography network is constructed, which includes a hiding network and a recovery network. Both the hiding network and the recovery network use an improved Swin Transformer as the core backbone, replacing the layer normalization module in the Swin Transformer with a dynamic Tanh normalization module. The audio steganography network was trained based on the training dataset to obtain the trained hidden network and the trained recovery network; the samples in the training dataset included host audio and secret audio. During training, the host audio and secret audio in the samples are converted into host temporal spectrograms and secret temporal spectrograms, respectively, and cascaded along the channel dimension as input to the hidden network. The hidden network outputs a container temporal spectrogram. The container temporal spectrogram is input to the recovery network to obtain the recovered secret temporal spectrogram. A discriminator network is then constructed, and the discriminator network performs adversarial training based on the host temporal spectrogram and the container temporal spectrogram output by the hidden network. The secret audio to be embedded and the selected host audio are converted into secret temporal spectrograms and host temporal spectrograms, respectively, and concatenated along the channel dimension. These are then input into the trained hidden network to obtain the container temporal spectrogram. An inverse short-time discrete cosine transform is performed on the container temporal spectrogram to obtain the container audio containing the secret information. Upon receiving a segment of container audio, a short-time discrete cosine transform is performed on the container audio to obtain the container time spectrum. The container time spectrum is then input into the trained recovery network to obtain the recovered secret time spectrum. Finally, an inverse short-time discrete cosine transform is performed on the recovered secret time spectrum to obtain the secret audio.

[0008] In one embodiment, the hidden network includes a shallow feature extraction module, a deep feature extraction module, a fusion module, and a temporal spectrogram reconstruction module; The shallow feature extraction module uses a 3x3 convolutional layer to perform preliminary feature mapping on the input cascaded time-spectrum map to obtain shallow features; The deep feature extraction module uses the Swin Transformer block to perform deep nonlinear transformations and information fusion on shallow features to obtain deep features. The fusion module is used to fuse deep features with shallow features to obtain preliminary fused features; then the preliminary fused features are fused with the cascaded time-spectrum map to obtain the final fused features; The temporal spectrogram reconstruction module includes a 1x1 convolutional layer and a 3x3 convolutional layer connected in sequence, which are used to perform two convolution operations on the final fused features to obtain the container temporal spectrogram.

[0009] In one embodiment, a Swing Transformer block includes: an image patch embedding layer, multiple cascaded dynamic residual Swing Transformer blocks, and an image patch de-embedding layer; The image patch embedding layer divides the input shallow features into multiple non-overlapping small image patches, and flattens and linearly projects each small image patch to obtain a high-dimensional vector. All high-dimensional vectors form a one-dimensional token sequence. The token sequence undergoes deep nonlinear transformation and information fusion through multiple cascaded dynamic residual Swing Transformer blocks to obtain the processed features. The processed feature input image block is de-embedded to restore a two-dimensional feature map, i.e., deep features.

[0010] In one embodiment, the dynamic residual Swin Transformer block includes a plurality of dynamic Swin Transformer layers, an image patch anti-embedding layer, and an image patch embedding layer connected in sequence. The dynamic Swin Transformer layer replaces the layer normalization module in the Swin Transformer layer with the dynamic Tanh normalization module. The input of the dynamic residual Swin Transformer block passes through multiple dynamic Swin Transformer layers, image patch anti-embedding layers, and image patch embedding layers, and the resulting output is fused with the input.

[0011] In one embodiment, the total loss used during training is:

[0012] in, For the total loss, Based on steganography loss, To help the generator combat loss, These are hyperparameters used to adjust the weights of the adversarial loss;

[0013]

[0014]

[0015] in, For host-container consistency loss, For secret-recovery consistency loss, Hyperparameters used to balance fidelity and recovery accuracy; The spectrum of the original host. To hide the spectrogram of the container generated by the network, The original secret time spectrum diagram, To recover the secret time-spectral graph generated by the network, This represents the square of the L2 norm.

[0016] Secondly, a high-fidelity audio steganography system based on an improved Swin Transformer is provided, including: The network building module is used to build audio steganography networks, which include a hiding network and a recovery network. Both the hiding network and the recovery network use an improved Swing Transformer as their core backbone, replacing the layer normalization module in Swing Transformer with a dynamic Tanh normalization module. The training module is used to train the audio steganography network based on the training dataset to obtain the trained hidden network and the trained recovery network; the samples in the training dataset include host audio and secret audio. During training, the host audio and secret audio in the samples are converted into host temporal spectrograms and secret temporal spectrograms, respectively, and cascaded along the channel dimension as input to the hidden network. The hidden network outputs a container temporal spectrogram. The container temporal spectrogram is input to the recovery network to obtain the recovered secret temporal spectrogram. A discriminator network is then constructed, and the discriminator network performs adversarial training based on the host temporal spectrogram and the container temporal spectrogram output by the hidden network. The secret information embedding module is used to convert the secret audio to be embedded and the selected host audio into secret temporal spectrograms and host temporal spectrograms respectively, and cascade them in the channel dimension. The concatenation is then input into the trained hidden network to obtain the container temporal spectrogram. The container temporal spectrogram is then subjected to inverse short-time discrete cosine transform to obtain the container audio containing the secret information. The secret information extraction module is used to perform a short-time discrete cosine transform on a received container audio segment to obtain a container time spectrum; input the container time spectrum into the trained recovery network to obtain the recovered secret time spectrum; and perform an inverse short-time discrete cosine transform on the recovered secret time spectrum to obtain the secret audio.

[0017] Compared with the prior art, this application has the following beneficial effects: 1. Innovative Network Architecture and Higher Embedding Quality. This application applies an improved Swin Transformer architecture (combined with dynamic Tanh normalization) to audio-to-audio steganography tasks. Compared to traditional convolutional neural networks (CNNs) or standard Transformers, this architecture can more effectively capture long-range time-frequency dependencies in the audio spectrogram, thereby achieving more refined and efficient information embedding and extraction. This allows this application to achieve both higher host audio fidelity (superior SDR, PESQ, and other metrics) and more accurate secret information recovery quality at the same embedding capacity.

[0018] 2. End-to-end optimization and enhanced concealment. This application constructs a complete end-to-end deep learning framework, avoiding the complex process of manually designing features and embedding rules required in traditional methods. By combining generative adversarial training (GAN), the model not only learns how to minimize the objective distortion of the signal, but also learns how to generate container audio that is perceptually and statistically difficult to distinguish from real audio, greatly improving the concealment and security of steganographic communication.

[0019] 3. Task-Specificity and Broad Application Prospects. This application focuses on solving the specific and challenging steganography task of "audio-to-audio". The method enables high-capacity secret audio embedding, providing a feasible technical solution for emerging application scenarios such as seamlessly embedding multilingual audio tracks in streaming content, adding hidden annotations in podcasts, and transmitting private commands to voice assistants. It has significant practical value and broad application prospects. Attached Figure Description

[0020] This application can be better understood by referring to the description given below in conjunction with the accompanying drawings, which, together with the detailed description below, are incorporated in and form part of this specification. In the drawings: Figure 1 A flowchart of a high-fidelity audio steganography method based on an improved Swing Transformer is shown. Figure 2 A schematic diagram of an audio steganography network is shown; Figure 3 A schematic diagram of a hidden network is shown; Figure 4 A schematic diagram of the dynamic residual Swing Transformer block is shown; Figure 5 A schematic diagram of the dynamic Swing Transformer layer is shown; Figure 6 A schematic diagram of the dynamic Tanh normalization module is shown. Figure 7 A spectral comparison diagram of the original host audio and the generated container audio is shown; Figure 8 A differential heatmap between the host audio and the container audio is shown. Detailed Implementation

[0021] Exemplary embodiments of the present application will be described below with reference to the accompanying drawings. For clarity and brevity, not all features of the actual embodiments are described in the specification. However, it should be understood that many embodiment-specific decisions can be made in the development of any such actual embodiment to achieve the developer’s specific objectives, and these decisions may vary as the embodiments differ.

[0022] It should also be noted that, in order to avoid obscuring this application with unnecessary details, only the device structure closely related to the solution of this application is shown in the accompanying drawings, while other details that are not closely related to this application are omitted.

[0023] It should be understood that this application is not limited to the described embodiments by virtue of the following description with reference to the accompanying drawings. In this document, embodiments may be combined with each other, features may be substituted or borrowed between different embodiments, and one or more features may be omitted in one embodiment, where feasible.

[0024] This application provides a high-fidelity audio steganography method based on an improved Swing Transformer. Figure 1 A flowchart of a high-fidelity audio steganography method based on an improved Swing Transformer is shown. See [link / reference]. Figure 1 The method mainly includes the following steps: Step S1: Construct an audio steganography network, which includes a hiding network and a recovery network. Figure 2 A schematic diagram of an audio steganography network is shown. Here, the hiding network and the recovery network have the same structure, and both networks use an improved Swin Transformer as their core backbone, replacing the layer normalization module in the Swin Transformer with a dynamic Tanh normalization module.

[0025] Step S2: Train the audio steganography network based on the training dataset to obtain the trained hidden network and the trained recovery network; the samples in the training dataset include host audio and secret audio.

[0026] During training, the host audio and secret audio in the samples are converted into host time-spectrum maps and secret time-spectrum maps, respectively, and cascaded in the channel dimension as input to the hidden network. The hidden network outputs a container time-spectrum map. The container time-spectrum map is input to the recovery network to obtain the recovered secret time-spectrum map. A discriminator network is constructed, and the discriminator network is trained adversarially based on the host time-spectrum map and the container time-spectrum map output by the hidden network.

[0027] Both the host audio and the secret audio in the sample are one-dimensional time-domain signals. To leverage the powerful two-dimensional data processing model, a Short-Time Discrete Cosine Transform (STDCT) module is first used to convert the two audio signals into two-dimensional host and secret time-domain spectrograms. These two time-domain spectrograms are cascaded along the channel dimension and fed into the core hidden network. Here, STDCT is chosen instead of the traditional Short-Time Fourier Transform (STFT) because its output is a real value, avoiding the complex phase estimation and reconstruction problems encountered when processing complex spectra. This allows the network model to focus on encoding and decoding amplitude information, simplifying model design and improving processing efficiency.

[0028] For example, the FSDNOsy18K dataset can be used as the host audio source, and the VCTK-Corpus dataset as the secret audio source. All audio samples undergo standardization preprocessing, including uniform resampling to the same sampling rate (e.g., 44.1 kHz) and normalizing the audio length to a fixed length (e.g., 67,522 samples, approximately 1.5 seconds) through truncation or padding to facilitate batch processing by the model. Next, a time-frequency transform is performed on the preprocessed time-domain audio signal. Using the Short-Time Discrete Cosine Transform (STDCT), with appropriate frame length, frame shift, and window function (e.g., Hamming window), each one-dimensional audio signal is converted into a two-dimensional, real-valued time-frequency spectrogram. For example, a time-frequency spectrogram with dimensions [512 × 128] can be generated. Finally, the generated time-frequency spectrogram is normalized, for example, by scaling its numerical range to the [-1, 1] interval to facilitate stable training of the neural network.

[0029] The hidden network is the core of information embedding; its task is to generate a container time-spectrum containing secret information, but which is visually highly similar to the host time-spectrum. To improve the imperceptibility of the generated spectrum, a discriminator network is introduced and trained adversarially against the hidden network. The discriminator receives the real host time-spectrum and the container time-spectrum generated by the hidden network and attempts to distinguish between the two.

[0030] Information extraction is the inverse process of embedding. The container time spectrogram is fed into the recovery network, which is responsible for decoding and recovering the secret time spectrogram.

[0031] Step S3: Convert the secret audio to be embedded and the selected host audio into secret time spectrogram and host time spectrogram respectively, and concatenate them in the channel dimension. Input them into the trained hidden network to obtain the container time spectrogram. Perform inverse short-time discrete cosine transform on the container time spectrogram to obtain the container audio containing secret information.

[0032] Step S4: After receiving a segment of container audio, perform a short-time discrete cosine transform on the container audio to obtain the container time spectrum; input the container time spectrum into the trained recovery network to obtain the recovered secret time spectrum; perform an inverse short-time discrete cosine transform on the recovered secret time spectrum to obtain the secret audio.

[0033] In the information embedding stage, the hidden network receives the concatenated host audio temporal spectrogram and secret audio temporal spectrogram as input. Through its deep, improved Swin Transformer-based coding structure, the network learns a complex nonlinear mapping that cleverly integrates the features of the secret audio into the feature representation of the host audio. The network output is a container audio temporal spectrogram of the same size as the host audio temporal spectrogram, which preserves both the main structure and energy distribution of the host audio and contains complete secret audio information. In the information recovery stage, the recovery network receives the container audio temporal spectrogram as input and, using its symmetric decoding structure, performs the inverse operation of the hidden network, separating and reconstructing the temporal spectrogram of the secret audio from the mixed features.

[0034] This embodiment maps the time-domain host audio and secret audio to a two-dimensional time-spectrum graph using Short-Time Discrete Cosine Transform (STDCT), which serves as the network input. A hidden network and a recovery network based on an improved Swin Transformer are designed to capture the multi-scale time-frequency dependencies in the time-spectrum graph. The time-spectrum graph of the secret audio is embedded into the time-spectrum graph of the host audio to generate container audio. A Generative Adversarial Network (GAN) mechanism is used to collaboratively optimize the entire framework, balancing the fidelity of the host audio, the recovery quality of the secret information, and the auditory imperceptibility. This embodiment systematically addresses the key challenges in audio-to-audio steganography tasks, designing a novel end-to-end deep learning framework that can achieve high-capacity, high-recovery-quality secret audio information hiding while maintaining high host audio fidelity, thus improving the concealment and reliability of steganographic communication.

[0035] In one embodiment, Figure 3 A schematic diagram of a hidden network is shown. The hidden network includes a shallow feature extraction module, a deep feature extraction module, a fusion module, and a temporal spectrogram reconstruction module. The shallow feature extraction module employs a 3x3 convolutional layer to perform preliminary feature mapping on the input cascaded spectrogram, obtaining shallow features. This module performs preliminary feature mapping on the original input spectrogram, capturing basic local patterns (such as harmonic structures and energy distributions), and expanding the number of channels in the feature map from 2 (host + secret) to the higher dimensions required for core network processing. This provides a richer and more standardized initial feature representation for subsequent deep modules.

[0036] The deep feature extraction module uses the Swin Transformer block to perform deep nonlinear transformations and information fusion on shallow features to obtain deep features. The fusion module is used to fuse deep features with shallow features to obtain preliminary fused features; then the preliminary fused features are fused with the cascaded time spectrogram to obtain the final fused features. Here, the fusion module is crucial, as it ensures that the network can simultaneously utilize fine shallow information and abstract deep semantics, so that the final generated container audio can retain the detailed structure of the host audio to the greatest extent and prevent information loss during transmission in the deep network.

[0037] The temporal spectrogram reconstruction module includes a 1x1 convolutional layer and a 3x3 convolutional layer connected in sequence, which are used to perform two convolution operations on the final fused features to obtain the container temporal spectrogram.

[0038] The feature map, after deep feature extraction and fusion, contains structural information of the host audio and encoded information of the secret audio. This feature map then enters the reconstruction stage. The first convolutional layer further integrates and refines the fused features, while the second convolutional layer is responsible for mapping the multi-channel feature map back to a single channel, generating the final container audio spectrogram. In this way, the network ensures that the generated container spectrogram contains both the secret information embedded by the deep network and maximizes the utilization of the preserved host-host audio structural information, thereby improving the fidelity and naturalness of the container audio.

[0039] Specifically, see Figure 3 The Swing Transformer block includes: Patch Embedding, multiple cascaded Dynamic Residual Swing Transformer blocks (D-RSTB), and Patch Unembedding. The image patch embedding layer divides the input shallow features into multiple non-overlapping small image patches, and flattens and linearly projects each small image patch to obtain a high-dimensional vector (token). All high-dimensional vectors form a one-dimensional token sequence, which is the standard input format that the Transformer model can handle. The token sequence undergoes deep nonlinear transformation and information fusion through multiple cascaded dynamic residual Swing Transformer blocks to obtain the processed features. The processed feature input image block is de-embedded to restore a two-dimensional feature map, i.e., deep features.

[0040] Specifically, Figure 4A schematic diagram of a dynamic residual Swing Transformer block is shown. The dynamic residual Swing Transformer block includes multiple dynamic Swing Transformer layers (D-STL), a patch unembedding layer, and a patch embedding layer connected in sequence. The dynamic Swing Transformer layer replaces the layer normalization module in the Swing Transformer layer with a dynamic Tanh normalization module. Figure 5 A schematic diagram of the dynamic Swing Transformer layer is shown. The dynamic Swing Transformer layer is the basic computational unit in this embodiment. It retains the core advantages of the standard Swing Transformer, namely, the alternating use of Window Multi-Head Self-Attention (W-MSA) and Shifted Window Multi-Head Self-Attention (SW-MSA) mechanisms. W-MSA restricts attention computation to non-overlapping local windows, significantly reducing computational complexity; while SW-MSA achieves cross-window information interaction through periodic window shifting, thereby capturing global context information.

[0041] The key improvement is to replace the original Layer Normalization module in D-STL with the Dynamic Tanh (DyT) Normalization module. Figure 6 A schematic diagram of the dynamic Tanh (DyT) normalization module is shown below. Figure 6 The mathematical form of the DyT module is γtanh(αx) + β, where x is the input feature, γ and β are learnable channel-level affine transformation parameters, and the core is a learnable scalar parameter α. This parameter α adaptively adjusts the activation scale of the tanh function, thereby achieving effective adjustment of the data distribution, especially compression of extreme values, without calculating runtime statistics (such as mean and variance). This design not only simplifies the computation process and improves inference efficiency, but also provides the network with a more flexible nonlinear transformation capability while maintaining or improving model performance stability, making it more suitable for processing audio temporal spectrograms and other data with a wide dynamic range.

[0042] The input of the dynamic residual Swin Transformer block passes through multiple dynamic Swin Transformer layers, image block anti-embedding layers, and image block embedding layers. The output is then fused with the input. This intra-block residual structure helps to build deeper networks and alleviate the gradient vanishing problem.

[0043] Specifically, in each training iteration, a batch of original host time-frequency spectrograms H and original secret time-frequency spectrograms S are randomly sampled from the prepared dataset. H and S are concatenated along the channel dimension and used as input to the hidden network. The hidden network performs forward computation and outputs a batch of container time-frequency spectrograms. .

[0044] Based on the results of the forward propagation, the basic steganalysis loss is calculated. The loss consists of two parts: Host-container consistency loss : Calculate the time spectrum H of the original host and the time spectrum of the generated container. The mean square error (MSE) between them, i.e. , This represents the square of the L2 norm. This loss term is designed to guarantee the fidelity of the container audio.

[0045] Secret - Recovery Consistency Loss : Spectrum diagram when calculating the original secret The Secret Time Spectrum of Recovery The mean square error between them, i.e. This loss item is designed to ensure the accuracy of the recovery of confidential information.

[0046] The weighted sum of the two yields the basic steganalysis loss: ,in It is a hyperparameter used to balance fidelity and recovery accuracy.

[0047] To further enhance the naturalness and imperceptibility of the container audio, making it statistically closer to the original host audio, this embodiment also introduces a Generative Adversarial Network (GAN) training mechanism. A discriminator network is designed to distinguish the spectrogram of the real host audio from the spectrogram of the "fake" container audio generated by the hidden network. During joint training, the hidden network (as a generator) not only minimizes the difference from the host audio and ensures the recoverability of secret information, but also strives to generate container audio that can "deceive" the discriminator. This adversarial game prompts the hidden network to learn more advanced and covert embedding strategies, resulting in container audio that is not only pixel-level similar but also difficult to identify in terms of auditory perception and data distribution. Adversarial training steps are performed simultaneously with the computation of the basic steganalysis loss.

[0048] First, the discriminator network is updated. A batch of original host time-spectrum maps H and container time-spectrum maps C generated by the hidden network are simultaneously input into the discriminator. The discriminator's goal is to correctly classify real samples and generated samples. Its loss function is... Standard adversarial loss is typically used, for example: Where D(·) represents the output probability of the discriminator. To calculate the mean, gradient descent is used to update only the parameters of the discriminator network, minimizing... This will enhance its ability to make judgments.

[0049] Then, update the generator (i.e., the hidden network). Calculate the generator's adversarial loss. The loss is designed to enable samples generated by the hidden network to "fool" the discriminator, that is, to allow... The value should be as close to 1 as possible. Its loss function can be defined as: .

[0050] Basic steganography loss Fighting loss against generators By performing a weighted summation, we obtain the total loss function used to update the hidden network and the recovery network. : ,in It is a hyperparameter used to adjust the weights of adversarial loss.

[0051] Calculate using the backpropagation algorithm The gradients relative to the hidden and recovered network parameters are calculated. Using a gradient descent optimizer (such as Adam), the model parameters of both the hidden and recovered networks are updated simultaneously based on the calculated gradients. Note that the discriminator parameters remain frozen during this step.

[0052] Repeat the above steps until the model's performance metrics (such as SDR, PESQ) on the validation set converge or meet the preset number of training rounds.

[0053] Employing the same inventive concept as the high-fidelity audio steganography method based on the improved Swing Transformer, this embodiment also provides a corresponding high-fidelity audio steganography system based on the improved Swing Transformer, including: The network building module is used to build audio steganography networks, which include a hiding network and a recovery network. Both the hiding network and the recovery network use an improved Swing Transformer as their core backbone, replacing the layer normalization module in Swing Transformer with a dynamic Tanh normalization module. The training module is used to train the audio steganography network based on the training dataset to obtain the trained hidden network and the trained recovery network; the samples in the training dataset include host audio and secret audio. During training, the host audio and secret audio in the samples are converted into host temporal spectrograms and secret temporal spectrograms, respectively, and cascaded along the channel dimension as input to the hidden network. The hidden network outputs a container temporal spectrogram. The container temporal spectrogram is input to the recovery network to obtain the recovered secret temporal spectrogram. A discriminator network is then constructed, and the discriminator network performs adversarial training based on the host temporal spectrogram and the container temporal spectrogram output by the hidden network. The secret information embedding module is used to convert the secret audio to be embedded and the selected host audio into secret temporal spectrograms and host temporal spectrograms respectively, and cascade them in the channel dimension. The concatenation is then input into the trained hidden network to obtain the container temporal spectrogram. The container temporal spectrogram is then subjected to inverse short-time discrete cosine transform to obtain the container audio containing the secret information. The secret information extraction module is used to perform a short-time discrete cosine transform on a received container audio segment to obtain a container time spectrum; input the container time spectrum into the trained recovery network to obtain the recovered secret time spectrum; and perform an inverse short-time discrete cosine transform on the recovered secret time spectrum to obtain the secret audio.

[0054] The high-fidelity audio steganography system based on the improved Swing Transformer in this embodiment has the same inventive concept as the high-fidelity audio steganography method based on the improved Swing Transformer described above. Therefore, the specific implementation of this device can be found in the embodiment section of the high-fidelity audio steganography method based on the improved Swing Transformer described above, and its technical effects correspond to the technical effects of the above method, so they will not be repeated here.

[0055] To verify the effectiveness and practical value of the end-to-end audio steganography method and system based on Swing Transformer proposed in this application, a systematic performance demonstration and analysis of its imperceptibility and robustness were conducted.

[0056] (1) Experimental data and evaluation indicators: performance analysis.

[0057] The FSDNoisy18K dataset was used as the host audio source, and the VCTK dataset as the secret audio source. To objectively evaluate the generation and reconstruction performance of the proposed method, signal-to-noise ratio (SNR), perceptual speech quality (PESQ), and short-time objective intelligibility (STOI) were used as core evaluation metrics. SNR and PESQ primarily measure the acoustic fidelity of the generated container audio compared to the original host audio; STOI measures the intelligibility and reconstruction quality of the secret audio extracted from the container audio.

[0058] (2) The demonstration of imperceptibility and qualitative effects.

[0059] The method described in this application can achieve high-capacity embedding of secret information while preserving the original characteristics of the host audio to a great extent. Figure 7 A spectral comparison diagram of the original host audio and the generated container audio is shown. Figure 7 As shown, the generated container audio maintains a high degree of consistency with the original host audio in terms of overall time-frequency structure and subtle frequency band texture. The difference is difficult to detect in both visual spectra and actual hearing, proving that the method of this application has excellent imperceptibility.

[0060] Figure 8 The differential heatmaps between host audio and container audio are shown, where (a) is the host time-frequency spectrum, (b) is the secret time-frequency spectrum, (c) is the container time-frequency spectrum, and (d) is the differential heatmap between host audio and container audio. Figure 8 As shown, the distribution of differential residual energy exhibits significant content-aware characteristics, with secret information adaptively hidden by the model in shadow regions of the host audio where energy is strong (such as strong harmonics or major speech components). This deep learning-based adaptive embedding strategy effectively utilizes the masking effect of the human auditory system, greatly improving the security of steganographic communication and effectively circumventing conventional steganalysis and detection.

[0061] (3) Demonstration of channel robustness and quantitative effect.

[0062] In real-world communication transmission scenarios, container audio inevitably suffers from various channel disturbances. To verify the engineering applicability of this application, the performance of the model was tested under clean conditions as well as with common channel disturbances such as additive white Gaussian noise (AWGN), audio resampling, and MP3 lossy compression. Specific robustness performance statistics are shown in Table 1.

[0063] Table 1. Robust performance statistics under common channel disturbances

[0064] Experimental data demonstrate that the proposed method maintains extremely high secret information recovery quality under undisturbed and low-to-moderate channel noise environments, with excellent STOI (Secret Transcription on Identification) metrics for the extracted secret speech. More importantly, even in complex communication environments with significant disruptions such as MP3 lossy compression and resampling, thanks to the powerful global time-frequency modeling capabilities of the improved Swin Transformer architecture, the proposed method can still effectively reconstruct secret features from damaged container audio, exhibiting outstanding anti-interference and robustness. This fully proves that the proposed method can meet the reliable communication requirements of real-world complex networks while ensuring high-fidelity information concealment.

[0065] The above descriptions are merely various embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A high-fidelity audio steganography method based on an improved Swing Transformer, characterized in that, include: An audio steganography network is constructed, which includes a hiding network and a recovery network. Both the hiding network and the recovery network use an improved Swing Transformer as the core backbone, replacing the layer normalization module in the Swing Transformer with a dynamic Tanh normalization module. The audio steganography network is trained based on the training dataset to obtain the trained hidden network and the trained recovery network; the samples in the training dataset include host audio and secret audio. During training, the host audio and secret audio in the samples are converted into host temporal spectrograms and secret temporal spectrograms, respectively, and cascaded along the channel dimension as input to the hidden network. The hidden network outputs a container temporal spectrogram. The container temporal spectrogram is input to the recovery network to obtain the recovered secret temporal spectrogram. A discriminator network is constructed, and the discriminator network performs adversarial training based on the host temporal spectrogram and the container temporal spectrogram output by the hidden network. The secret audio to be embedded and the selected host audio are converted into secret temporal spectrograms and host temporal spectrograms, respectively, and concatenated in the channel dimension. These are then input into the trained hidden network to obtain the container temporal spectrogram. An inverse short-time discrete cosine transform is performed on the container temporal spectrogram to obtain the container audio containing the secret information. Upon receiving a segment of container audio, a short-time discrete cosine transform is performed on the container audio to obtain a container time-spectrum; the container time-spectrum is input into the trained recovery network to obtain a recovered secret time-spectrum; an inverse short-time discrete cosine transform is performed on the recovered secret time-spectrum to obtain the secret audio.

2. The method as described in claim 1, characterized in that, The hidden network includes a shallow feature extraction module, a deep feature extraction module, a fusion module, and a time-spectrum reconstruction module; The shallow feature extraction module uses a 3x3 convolutional layer to perform preliminary feature mapping on the input cascaded time-spectrum map to obtain shallow features; The deep feature extraction module uses the Swing Transformer block to perform deep nonlinear transformation and information fusion on the shallow features to obtain deep features. The fusion module is used to fuse the deep features with the shallow features to obtain preliminary fused features; then, the preliminary fused features are fused with the cascaded time-spectrum map to obtain the final fused features; The temporal spectrogram reconstruction module includes a 1x1 convolutional layer and a 3x3 convolutional layer connected in sequence, which are used to perform two convolution operations on the final fused features to obtain the container temporal spectrogram.

3. The method as described in claim 2, characterized in that, The Swin Transformer block includes: an image block embedding layer, multiple cascaded dynamic residual Swin Transformer blocks, and an image block de-embedding layer; The image patch embedding layer divides the input shallow features into multiple non-overlapping small image patches, and flattens and linearly projects each small image patch to obtain a high-dimensional vector. All high-dimensional vectors form a one-dimensional token sequence. The token sequence undergoes deep nonlinear transformation and information fusion through multiple cascaded dynamic residual Swing Transformer blocks to obtain processed features. The processed features are input into the image block anti-embedding layer to restore a two-dimensional feature map, i.e., deep features.

4. The method as described in claim 3, characterized in that, The dynamic residual Swin Transformer block includes multiple dynamic Swin Transformer layers, an image block anti-embedding layer, and an image block embedding layer connected in sequence. The dynamic Swin Transformer layer replaces the layer normalization module in the Swin Transformer layer with a dynamic Tanh normalization module. The input of the dynamic residual Swing Transformer block passes through multiple dynamic Swing Transformer layers, image block anti-embedding layers, and image block embedding layers, and the resulting output is fused with the input.

5. The method as described in claim 1, characterized in that, The total loss used during training is: in, For the total loss, Based on steganography loss, To help the generator combat loss, These are hyperparameters used to adjust the weights of the adversarial loss; in, For host-container consistency loss, For secret-recovery consistency loss, Hyperparameters used to balance fidelity and recovery accuracy; The spectrum of the original host. To hide the spectrogram of the container generated by the network, The original secret time spectrum diagram, To recover the secret time-spectral graph generated by the network, This represents the square of the L2 norm.

6. A high-fidelity audio steganography system based on an improved Swing Transformer, characterized in that, include: A network construction module is used to construct an audio steganography network, which includes a hiding network and a recovery network. Both the hiding network and the recovery network use an improved Swin Transformer as their core backbone, replacing the layer normalization module in the Swin Transformer with a dynamic Tanh normalization module. The training module is used to train the audio steganography network based on the training dataset to obtain the trained hidden network and the trained recovery network; the samples in the training dataset include host audio and secret audio. During training, the host audio and secret audio in the samples are converted into host temporal spectrograms and secret temporal spectrograms, respectively, and cascaded along the channel dimension as input to the hidden network. The hidden network outputs a container temporal spectrogram. The container temporal spectrogram is input to the recovery network to obtain the recovered secret temporal spectrogram. A discriminator network is constructed, and the discriminator network performs adversarial training based on the host temporal spectrogram and the container temporal spectrogram output by the hidden network. The secret information embedding module is used to convert the secret audio to be embedded and the selected host audio into secret time-spectrum maps and host time-spectrum maps respectively, and cascade them in the channel dimension. The concatenation is then input into the trained hidden network to obtain the container time-spectrum map. The container time-spectrum map is then subjected to inverse short-time discrete cosine transform to obtain the container audio containing the secret information. The secret information extraction module is used to perform a short-time discrete cosine transform on the received container audio to obtain a container time spectrum; input the container time spectrum into the trained recovery network to obtain a recovered secret time spectrum; and perform an inverse short-time discrete cosine transform on the recovered secret time spectrum to obtain the secret audio.

Images