Steganographic image detection and cleaning method

Abstract

The application discloses a steganographic image detection and purification method, and constructs a steganographic defense model comprising a steganographic analysis module and a flexible erasing module. The flexible erasing module is realized based on a conditional resampling technology, wherein a scale parameter is dynamically set according to a steganographic confidence score of a current input image detected by the steganographic analysis module. The steganographic analysis module and the flexible erasing module are trained respectively, and the trained steganographic defense model is used for detecting and purifying the input image. The steganographic image is efficiently detected and purified through the flexible erasing module and the dynamic setting of the scale parameter according to the steganographic confidence score.

Description

Technical Field

[0001] This invention belongs to the field of information security technology, and more specifically, relates to a method for detecting and cleaning up steganalog images. Background Technology

[0002] Image steganography is an information hiding technique that embeds secret information into a carrier image, making this information visually invisible, thereby achieving covert transmission or storage. Steganography is widely used in fields such as digital watermarking, information protection, and privacy communication. However, with the development of digital information and network technologies, this technology has gradually attracted attention to information security, especially in its application to protect sensitive information and prevent information leakage. In recent years, the rapid advancements in deep learning have played a transformative role in image steganography research, making the implementation methods of steganography more complex and intelligent. Deep learning models, particularly excelling in feature extraction and information embedding, have significantly improved the robustness and concealment of steganographic information, making steganographic images more difficult to detect and decipher.

[0003] The double-edged nature of this technology means that while it provides legitimate users with a means of privacy protection, it also provides potential tools for cybercrime and information theft, threatening public safety. Especially in contexts involving national security, corporate secrets, and personal privacy, the stealth and complexity of deep learning-driven steganography renders traditional detection and protection methods inadequate.

[0004] Traditional steganalysis methods rely on manually designed feature extraction techniques combined with machine learning classifiers to detect and identify steganographic information. While these methods achieved some success in their early stages, they have become less effective against deep learning-generated steganography with the widespread application of deep learning technology. Deep learning-driven steganography, by automatically learning hidden patterns and complex features in data, overcomes the limitations of manual feature extraction, making the detection and deciphering of steganographic information more difficult and inefficient.

[0005] To address the challenges of deep learning steganography, active steganalysis methods have emerged. These methods attempt to disrupt steganalytic information and undermine its effectiveness by preprocessing images, such as image flipping, adding noise, and compression loss. However, these perturbation methods typically have a significant impact on image quality, reducing visibility and clarity, thus affecting user experience and even making the image difficult to use in practice. While these methods improve the likelihood of steganalytic information detection to some extent, their side effects and limitations make them difficult to widely apply in real-world scenarios.

[0006] To address these challenges, current steganalysis (SD) defense techniques primarily employ image processing methods to remove or reduce steganographic information while preserving image visual quality as much as possible. Typical defense techniques include denoising, compression, and image reconstruction. However, these methods often struggle to achieve an ideal balance between the effectiveness of steganographic removal and the preservation of image quality. Especially when the steganographic removal rate is increased, image quality is often significantly affected, leading to visual distortion or information loss, which limits their effectiveness in real-world applications. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for detecting and cleaning up stegana images. By setting up a flexible erasure module and dynamically setting scale parameters according to the stegana confidence score, efficient detection and cleaning of stegana images can be achieved.

[0008] To achieve the above-mentioned objectives, the method for detecting and cleaning up stegographic images of the present invention includes the following steps:

[0009] S1: Construct a steganography defense model, including a steganography analysis module and a flexible erasure module, wherein:

[0010] The steganalysis module is used to analyze the input image I in Steganography detection is performed to obtain a steganalysis confidence score s. If the steganalysis confidence score s > S, where S represents a preset threshold, then the input image I is considered safe. in To create a steganographic image, input image I... in The steganography confidence score s is sent to the flexible erase module; otherwise, no action is taken.

[0011] The flexible erasure module is used to remove the input image I based on the steganalysis confidence score s. in Stegature information in the image is used to clean up the image I. eas The flexible erase module includes a downsampling module and an upsampling module, wherein:

[0012] The downsampling module is used to process the input image I in Downsampling is performed to obtain the intermediate image I. temp And send it to the upsampling module; the downsampling module includes an input image feature extraction module, a conditional resampling module, a third convolutional layer, and a quantization module, wherein:

[0013] The input image feature extraction module is used to extract features from the input image I. in Feature extraction is performed, and the obtained features are sent to the conditional resampling module;

[0014] The conditional resampling module is used to downsample the received features according to the scale parameter ε using a conditional resampling mechanism, where the scale parameter... ω represents the preset increment parameter, and its value range is ω∈[1,2]. This indicates that the integer part is rounded up and the resulting features are sent to the third convolutional layer.

[0015] The third convolutional layer is used to perform convolution operations on the received features and send the resulting features to the quantization layer.

[0016] The quantization layer is used to quantize the received features to obtain the intermediate image I. temp ;

[0017] The upsampling module is used for the intermediate image I temp Upsampling is performed to obtain the cleaned image I. eas The upsampling module includes an intermediate image feature extraction module, a conditional resampling module, and a third convolutional layer, wherein:

[0018] The intermediate image feature extraction module is used to extract features from intermediate image I. temp Feature extraction is performed, and the obtained features are sent to the conditional resampling module;

[0019] The conditional resampling module is used to upsample the received features according to the scale parameter ε using a conditional resampling mechanism, and then send the obtained features to the third convolutional layer.

[0020] The third convolutional layer performs a convolution operation on the received features to obtain the cleaned image I. eas ;

[0021] S2: Collect several steganalytical images and clean images as training samples to train the steganalysis module and obtain a trained steganalysis module.

[0022] S3: Collect several steganalysis images according to actual needs, randomly generate steganalysis confidence scores for each steganalysis image, and use them as training samples to train the flexible erasure module to obtain a trained steganalysis module.

[0023] S4: The image I needs to be detected and cleaned i ′ n Input the trained steganalysis defense model to obtain the steganalysis confidence score s′. When the steganalysis confidence score s′>S, the cleaned-up image I is obtained. e ′ as .

[0024] The present invention provides a method for detecting and cleaning stegographic images. It constructs a steganalysis module and a flexible erasure module, wherein the flexible erasure module is implemented based on conditional resampling technology. The scale parameter is dynamically set according to the steganalysis confidence score obtained by the steganalysis module from the current input image. The steganalysis module and the flexible erasure module are trained respectively, and the trained steganalysis module is used to detect and clean the input image.

[0025] The present invention has the following beneficial effects:

[0026] 1) This invention allows users to find a suitable balance between steganalysis removal rate and image quality based on specific security requirements by dynamically adjusting the scale parameters. Different steganalysis strategies in different scenarios can be implemented by adjusting the parameters, which provides users with great flexibility.

[0027] 2) This invention utilizes a pyramid attention mechanism to enhance feature representation, enabling the model to maintain a high level of detection accuracy when processing complex images. Furthermore, the Conditional Resampling (CRM) module endows the framework with upsampling and downsampling capabilities, effectively controlling the fine preservation of image information, reducing visual distortion caused by erasure, improving the quality of cleaned images, and meeting the balance between high steganalytic removal rate and good visual effects.

[0028] 3) In experimental verification under various datasets and steganography methods, the present invention has demonstrated excellent performance, especially in terms of steganography removal rate and image similarity. The present invention can effectively remove hidden information while maintaining high image quality, proving its reliability and effectiveness in practical applications. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating a specific implementation method for the steganography detection and cleanup method of the present invention;

[0030] Figure 2 This is a structural diagram of the steganography analysis module in this embodiment;

[0031] Figure 3 This is a structural diagram of the flexible erasure module in this invention;

[0032] Figure 4 This is a graph showing the change in damage rate as a function of scale parameters in this embodiment;

[0033] Figure 5 This is a comparison chart of the steganalysis performance of the present invention and two comparative methods against robust steganalysis algorithms in this embodiment;

[0034] Figure 6This is a comparison chart of the steganalysis performance of the present invention and the comparison method against non-robust steganalysis algorithms in this embodiment. Detailed Implementation

[0035] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0036] Example

[0037] Figure 1 This is a flowchart illustrating a specific implementation of the steganography detection and cleanup method of the present invention. For example... Figure 1 As shown, the specific steps of the steganography detection and cleanup method of the present invention include:

[0038] S101: Constructing a Steganography Defense Model:

[0039] To efficiently clean up stegated images, this invention constructs a steg defense model (ErastegNet). This steg defense model includes a steganalyzer module and a flexible erasure module, which will be described in detail below.

[0040] The steganalysis module is a fundamental component of ErastegNet, used to perform steganalysis on the input image and obtain a steganalysis confidence score s. If the steganalysis confidence score s > S, where S represents a preset threshold, then the input image I is considered steganalysis. in For steganographic images, the input image and steganographic confidence score s are sent to the flexible erasure module; otherwise, no operation is performed. In this embodiment, the threshold S = 0.5.

[0041] In this embodiment, in order to improve the detection accuracy of stegographic images, a steganalysis module is constructed by combining multiple steganalysis methods. Figure 2 This is a structural diagram of the steganography analysis module in this embodiment. For example... Figure 2 As shown, in this embodiment, the steganalysis module includes K steganalysis detection modules and a detection result fusion module, wherein:

[0042] Each steganalysis module employs a different steganalysis method to perform steganalysis on the input image, obtaining the corresponding steganalysis confidence score. Then the steganalysis confidence score s kThe results are sent to the detection result fusion module. In this embodiment, the number of steganalysis modules is K=3, namely, a steganalysis module based on SRNet, a steganalysis module based on YedroudjNet, and a steganalysis module based on SiaStegNet. These three steganalysis modules have their own advantages in different feature extraction and detection, enabling the steganalysis module to detect steganalysis information in the input image from multiple dimensions.

[0043] The detection result fusion module is used to process K stegwriting confidence scores. The steganalysis is performed to obtain a steganalysis confidence score s. Each steganalysis confidence score represents the confidence level of the corresponding steganalysis detection module in whether the input image contains steganalysis information. The final steganalysis confidence score s is obtained by fusing K steganalysis confidence scores. In this embodiment, the steganalysis confidence score s is K steganalysis confidence scores. The maximum value in, i.e.

[0044] The flexible erasure module is the core processing unit of ErastegNet, used to remove steg information from the input image based on the stegacy confidence score s to obtain a cleaned image. Figure 3 This is a structural diagram of the flexible erasure module in this invention. (See diagram for example.) Figure 3 As shown, the flexible erasure module in this invention includes a downsampling module and an upsampling module. The two modules will be described in detail below.

[0045] The downsampling module is used to process the input image I in Downsampling is performed to obtain the intermediate image I. temp The image is then sent to the upsampling module. The downsampling module includes an input image feature extraction module, a conditional resampling module (CRM), a third convolutional layer, and a quantization module, wherein:

[0046] The input image feature extraction module is used to extract features from the input image I. in Feature extraction is performed, and the obtained features are sent to the conditional resampling module.

[0047] The conditional resampling module is used to downsample the received features according to the scale parameter ε using a conditional resampling mechanism, where the scale parameter... ω represents the preset increment parameter, and its value range is ω∈[1,2]. This indicates that the data is rounded up and the resulting features are sent to the third convolutional layer. By dynamically calculating the scale parameter, accurate detection and processing under different steganalysis intensities are ensured.

[0048] The conditional resampling module is a dynamic resampling module designed to flexibly handle image scaling requirements. Unlike traditional fixed-ratio (e.g., 2x, 4x) interpolation methods, the conditional resampling module can adaptively scale to any ratio (e.g., 1.3x or 0.7x) and dynamically adjust the sampling strategy based on image content. Its core lies in combining scale parameters with local image features to automatically generate a dynamic sampling kernel and coordinate offset, handling both upsampling and downsampling. This design breaks away from the dependence of traditional methods on integer scaling multiples, while preserving high-frequency details (such as edges and textures) through a content-aware mechanism, avoiding blurring or jagged edges caused by conventional interpolation. In the field of steganography security, conditional resampling effectively eliminates hidden data by dynamically disrupting the positional correlation of steganographic information, while maintaining image visual quality. For the specific structure and working principle of the conditional resampling module, please refer to the literature "Xing J, Hu W, Wong T T. Scale-arbitrary Invertible ImageDownscaling[J]. 2022. DOI:10.48550 / arXiv.2201.12576."

[0049] The third convolutional layer is used to perform convolution operations on the received features and send the resulting features to the quantization layer.

[0050] The quantization layer is used to quantize the received features to obtain the intermediate image I. temp .

[0051] The upsampling module is used for the intermediate image I temp Upsampling is performed to obtain the cleaned image I. eas The structure of the upsampling module is similar to that of the downsampling module, including an intermediate image feature extraction module, a conditional resampling module, and a third convolutional layer, wherein:

[0052] The intermediate image feature extraction module is used to extract features from intermediate image I. temp Feature extraction is performed, and the obtained features are sent to the conditional resampling module.

[0053] The conditional resampling module is used to upsample the received features according to the scale parameter ε using a conditional resampling mechanism, and then send the resulting features to the third convolutional layer.

[0054] The third convolutional layer performs a convolution operation on the received features to obtain the cleaned image I. eas .

[0055] like Figure 3As shown, in this embodiment, the input image feature extraction module and the intermediate image feature extraction module adopt the same structure, which respectively include a first convolutional layer (Conv), a first residual module (Res), a pyramid attention module (PyramidAttention), a second residual module, a second convolutional layer, and a feature fusion module, wherein:

[0056] The first convolutional layer is used to process the input image I. in Perform convolution operations and output the resulting features to the first residual module and the feature fusion module.

[0057] The first residual module is used to process the received features using a residual mechanism and output the obtained features to the pyramid attention module.

[0058] The pyramid attention module extracts features from the received features using the pyramid attention mechanism and sends the extracted features to the second residual module. The pyramid attention mechanism is a technique in deep learning used to improve the performance of convolutional neural networks. It utilizes multi-scale input feature maps to extract and integrate spatial information at different scales, thereby establishing long-term dependencies between multi-scale channel attention, thus improving the accuracy and fine-grained focus of feature extraction. By weighting information at different scales, the pyramid attention mechanism helps the model focus its attention on key regions, enhancing the effectiveness of steganalysis detection and removal.

[0059] The second residual module is used to process the received features using a residual mechanism and output the obtained features to the second convolutional layer.

[0060] The second convolutional layer is used to perform convolution operations on the received features and send the resulting features to the feature fusion module.

[0061] The feature fusion module is used to superimpose two received features and output the resulting fused feature.

[0062] S102: Training the steganalysis module:

[0063] Collect several steganalytical images and clean images as training samples to train the steganalysis module, thus obtaining a trained steganalysis module.

[0064] Since the steganalysis module is a binary classification model, the loss function used in this embodiment is cross-entropy loss.

[0065] S103: Training the flexible erase module:

[0066] Collect several steganases according to actual needs, randomly generate steganasis confidence scores for each steganases, and use them as training samples to train the flexible erasure module, thus obtaining a trained flexible erasure module.

[0067] In this embodiment, the loss function LOSS during the training of the flexible erasure module is divided into two parts: intermediate image loss L... ref and purification loss L eas Intermediate image loss L ref The calculation method is as follows:

[0068] For input image I in The reference intermediate image I is obtained by direct downsampling based on the scale parameter ε. ref In this embodiment, direct downsampling uses a bicubic interpolation method. The intermediate image loss is L. ref Used to measure intermediate image I temp With reference image L ref The difference between them; then the intermediate image loss L is calculated using the following formula. ref :

[0069] L ref =||I temp -I ref ||2

[0070] Where || ||2 represents finding the L2 norm.

[0071] Purification loss L eas Intended to measure image purification I eas With input image I in The difference between them is calculated using the following formula:

[0072] L eas =||I in -I eas ||1

[0073] Here, || ||1 represents finding the first norm.

[0074] Then, the loss function LOSS is calculated using the following formula:

[0075] LOSS=λ×L eas +L ref

[0076] Where λ represents the preset weight.

[0077] S104: Steganography Detection and Cleaning:

[0078] The image I that needs to be detected and cleaned i ′ nInput the trained steganalysis defense model to obtain the steganalysis confidence score s′. When the steganalysis confidence score s′>S, the cleaned-up image I is obtained. e ′ as .

[0079] To better illustrate the technical solution of the present invention, specific examples are used to experimentally verify the present invention.

[0080] In this embodiment, two grayscale image datasets (BOSS1.0.1 and ALASKA) and two color image datasets (ImageNet and CelebA) were used to train the steganalysis module. The steganalytes generated using these datasets cover a variety of steganalysis algorithms, including two traditional steganalysis methods, HUGO and S-UNIWARD (0.4 bpp payload), two deep learning-based steganalysis methods, HiNet and Weng (256×256 image size), and two robust steganalysis methods, Rosteals and StegaStamp (100-bit payload). Ultimately, 360,000 clean / steganographic image pairs (color and grayscale) were generated to train the steganalysis module, improving its versatility on online social platforms.

[0081] This embodiment implements the steganalysis defense model (ErastegNet) on Ubuntu 18.04 using PyTorch 2.0. It was trained for 400 epochs on an NVIDIA A100 GPU using the Adam optimizer. All input images were resized to 256×256. The increment parameter ω for the scale parameter ε was 1.5.

[0082] To verify the impact of the flexible scale parameter ε on steganalysis, the Destruction Rate (DR) was tested under different scale parameters ε. First, 12,000 256×256 steganalytes (3,000 per dataset) were generated on the CIFAR-10, BOSS 1.0.1, and ImageNet datasets using the ISGAN steganography method. These steganalytes were then fed into the flexible erasure module of this invention to obtain the cleaned images, and the DR value was calculated. Figure 4 This is a graph showing the damage rate as a function of scale parameters in this embodiment. For example... Figure 4 As shown, the DR value gradually increases with the increase of the scale parameter ε, but the growth of the DR value slows down significantly when the scale parameter ε ≥ 1.6. Therefore, the damage rate can be controlled more effectively through dynamic scale parameters.

[0083] This embodiment employs two of the best existing steganalysis methods for detection and cleanup—PixelSteganalysis—as comparison methods, along with Gaussian noise and scaling methods (denoted as Gaussiannoise). In this embodiment, DEEP and ISGAN are used to generate steganalytes, which are then cleaned using the present invention and the two comparison methods. Each method is used to generate 1000 images from ImageNet, CIFAR-10, and BOSS 1.0.1 datasets. The evaluation metric used in this embodiment is image similarity (PSNR), which refers to the similarity between the cleaned image and the original image; a higher PSNR indicates less image quality loss.

[0084] Figure 5 This is a comparison chart of the steganalysis performance of the present invention and two comparative methods against robust steganography algorithms in this embodiment. Figure 5 As shown, the present invention performs excellently against robust steganography algorithms DMAS and GMAS, achieving the highest PSNR value, indicating that it can maintain image quality to a high degree while removing hidden data.

[0085] To test the performance of this invention against non-robust steganography algorithms, this embodiment uses three datasets to conduct comparative experiments on this invention and two comparative methods. Figure 6 This is a comparison chart of the steganalysis performance of the present invention and the comparison method against robust steganalysis algorithms in this embodiment. For example... Figure 6 As shown, the algorithm of this invention also demonstrates superior performance against non-robust steganography.

[0086] In summary, the present invention outperforms existing traditional methods in all performance indicators of steganalysis and image cleansing. Compared with these methods, the present invention has significant advantages in steganalysis removal rate and image quality preservation, fully demonstrating the innovation of the present invention and its potential in practical applications.

[0087] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A method for detecting and cleaning up steganalog images, characterized in that, Includes the following steps: S1: Construct a steganography defense model, including a steganography analysis module and a flexible erasure module, wherein: The steganalysis module is used to analyze the input image. Perform steganalysis and obtain steganalysis confidence scores. When the steganographic confidence score , This indicates the preset threshold, which is used to determine the input image. To create a steganographic image, the input image is... and steganalysis confidence score Send to the flexible erase module; otherwise, do nothing. The flexible erase module is used to determine the steganography confidence score. Remove input image Stegature information in the image is purified. The flexible erase module includes a downsampling module and an upsampling module, wherein: The downsampling module is used to process the input image. Downsampling is performed to obtain the intermediate image. And send it to the upsampling module; the downsampling module includes an input image feature extraction module, a conditional resampling module, a third convolutional layer, and a quantization module, wherein: The input image feature extraction module is used to process the input image. Feature extraction is performed, and the obtained features are sent to the conditional resampling module; The conditional resampling module is used to determine the scale parameter. A conditional resampling mechanism is used to downsample the received features, where the scale parameter... , This represents the preset incremental parameter, and its value range is... , This indicates that the integer part is rounded up and the resulting features are sent to the third convolutional layer. The third convolutional layer is used to perform convolution operations on the received features and send the resulting features to the quantization layer. The quantization layer is used to quantize the received features to obtain an intermediate image. ; The upsampling module is used for intermediate images Upsampling is performed to obtain a cleaned image. The upsampling module includes an intermediate image feature extraction module, a conditional resampling module, and a third convolutional layer, wherein: The intermediate image feature extraction module is used to extract features from intermediate images. Feature extraction is performed, and the obtained features are sent to the conditional resampling module; The conditional resampling module is used to determine the scale parameter. The received features are upsampled using a conditional resampling mechanism, and the resulting features are sent to the third convolutional layer. The third convolutional layer performs convolution operations on the received features to obtain the cleaned image. ; S2: Collect several steganalytical images and clean images as training samples to train the steganalysis module and obtain a trained steganalysis module. S3: Collect several stegana images according to actual needs, randomly generate steganalysis confidence scores for each stegana image, and use them as training samples to train the flexible erasure module to obtain a trained flexible erasure module. S4: Images that need to be inspected and cleaned Input the trained steganalysis defense model and obtain the steganalysis confidence score. When the steganographic confidence score At that time, a purified image was obtained. .

2. The steganalysis detection and cleanup method according to claim 1, characterized in that, The steganography analysis module includes A steganography detection module and a detection result fusion module, wherein: Each steganalysis module employs a different steganalysis method to analyze the input image. Perform steganalysis and obtain the corresponding steganalysis confidence score. , Then, steg confidence scores Send to the detection result fusion module; The detection result fusion module is used to... 1 stegographic confidence score The fusion is performed to obtain the steganalysis confidence score. .

3. The method for detecting and cleaning steganalog images according to claim 2, characterized in that, The stegographic confidence score for 1 stegographic confidence score The maximum value in.

4. The method for detecting and cleaning steganalog images according to claim 1, characterized in that, The threshold .

5. The method for detecting and cleaning steganalog images according to claim 1, characterized in that, The input image feature extraction module and the intermediate image feature extraction module adopt the same structure, each including a first convolutional layer, a first residual module, a pyramid attention module, a second residual module, a second convolutional layer, and a feature fusion module, wherein: The first convolutional layer is used to process the input image. Perform convolution operations and output the resulting features to the first residual module and the feature fusion module; The first residual module is used to process the received features using a residual mechanism and output the obtained features to the pyramid attention module. The pyramid attention module is used to extract features from the received features using the pyramid attention mechanism, and the obtained features are sent to the second residual module. The second residual module is used to process the received features using a residual mechanism and output the obtained features to the second convolutional layer. The second convolutional layer is used to perform convolution operations on the received features and send the resulting features to the feature fusion module. The feature fusion module is used to superimpose two received features and output the resulting fused feature.

6. The method for detecting and cleaning stegographic images according to claim 1, characterized in that, The loss function used for training the steganalysis module is cross-entropy loss.

7. The method for detecting and cleaning steganalog images according to claim 1, characterized in that, The loss function during training of the flexible erasure module The calculation method is as follows: For the input image According to scale parameters Direct downsampling is performed to obtain the reference intermediate image. Then, the intermediate image loss is calculated using the following formula. : ， in, This indicates the search for the L2 norm; The purification loss is calculated using the following formula. : ， in, This indicates the search for a norm. Then, the loss function is calculated using the following formula. : ， in, This indicates the preset weight.

Images