Frequency domain multi-target information steganography technology based on discrete wavelet transform

Abstract

The invention provides a frequency domain multi-target information steganography technology based on discrete wavelet transform, and relates to the field of deep neural network security, and the technology specifically comprises the following steps: converting an RGB color image into a YUV space, and extracting a UV chromaticity channel; a UV channel of the image is transformed to a frequency domain through DCT, and an information trigger is embedded in a high-frequency area; using the feature extraction capability of discrete wavelet transform to take the four types of features as four targets; and converting the YUV space back to the RGB space to obtain a finally processed image, and then training the image in a neural network model to evaluate the effectiveness and concealment of the information steganography technology. According to the information steganography technology, the problems that in the prior art, information steganography is low in effectiveness and concealment and limited to a single target can be solved.

Description

Technical Field

[0001] This invention relates to the field of deep neural network security, and to the fields of deep learning and digital image processing technology. Specifically, it relates to a frequency domain multi-target information steganography technique based on discrete wavelet transform. Background Technology

[0002] Deep neural networks (DNNs) have achieved remarkable success in various fields such as image classification, speech recognition, and natural language processing, and are widely used in many real-world scenarios, such as autonomous driving, medical image analysis, and facial recognition. Steganography is a method of embedding secret information into carriers such as images, audio, and text, making it "invisible" to the naked eye or conventional detection. This allows the recipient to extract the hidden content without loss of quality, thus achieving covert transmission and providing "traceless" communication methods for military, intelligence, and commercial fields. Processors implant specific information triggers into the training data. These triggers can be activated under specific conditions, allowing the model to obtain the specified information from the input with the trigger during the inference phase, while maintaining the original results on unprocessed samples.

[0003] Numerous studies have proposed diverse steganography techniques to enhance the effectiveness and concealment of information. Most existing steganography methods focus on improving the invisibility of information triggers in the spatial domain, such as early methods like BadNets and Blend. However, these are easily detectable. Current research on designing concealed information triggers in the frequency domain is relatively limited, and these triggers are easily detected by frequency-based detection methods. Furthermore, most of these studies address a single objective, leaving room for further improvement in the effectiveness of information steganography.

[0004] The masking characteristics of the human visual system (HVS) typically include: frequency characteristics that are less sensitive to high-frequency content than low-frequency content; color characteristics that are less sensitive to chromaticity than to luminance; phase characteristics that are less sensitive to phase changes than to amplitude changes; and motion characteristics that are less sensitive to the spatial characteristics of moving objects than to the characteristics of stationary objects.

[0005] To address the issue that the concealment and effectiveness of current information steganography methods in the frequency domain need to be optimized, this paper proposes a frequency domain multi-target information steganography technique based on discrete wavelet transform. This technique is based on the frequency characteristics of the human visual system (HVS), which are less sensitive to high frequencies than low frequencies, and the color characteristics, which are less sensitive to chromaticity than luminance. It also utilizes the feature extraction capabilities of discrete wavelet transform and proposes a technique through color space transformation and frequency domain space transformation. This technique can effectively improve the concealment and processing effectiveness of information steganography. Summary of the Invention

[0006] The main objective of this invention is to provide a frequency domain multi-target information steganography technique based on discrete wavelet transform, so as to solve the problem of low concealment and effectiveness of existing processing methods.

[0007] To achieve the above objectives, this invention provides a frequency domain multi-target information steganography technique based on discrete wavelet transform, specifically including the following steps:

[0008] (a) Convert the RGB color image to YUV space and extract the UV chromaticity channels;

[0009] (ii) Transform the UV channels of the image to the frequency domain using DCT and embed information triggers in the high-frequency region;

[0010] (iii) Utilizing the feature extraction capability of discrete wavelet transform, the four types of features are used as four targets;

[0011] (iv) Convert the YUV space back to the RGB space to obtain the final processed image, and then train it in a neural network model to evaluate the effectiveness and concealment of the information steganography method.

[0012] Furthermore, step (one) specifically includes the following:

[0013] Step (1) Convert the RGB color image to the YUV color space using the following formula:

[0014] Y = 0.299 * R + 0.587 * G + 0.114 * B,

[0015] U = 0.596 * R - 0.272 * G - 0.321 * B,

[0016] V = 0.212 * R - 0.523 * G - 0.311 * B.

[0017] Furthermore, step (two) specifically includes the following:

[0018] Step (ii) transforms the UV channel of the image to the frequency domain using DCT and embeds the information trigger into the high-frequency region based on the characteristic that the human visual system is less sensitive to high-frequency components than to low-frequency components.

[0019] Furthermore, step (iii) specifically includes the following:

[0020] Step (3) utilizes the feature extraction capability of Discrete Wavelet Transform (DWT) to treat four types of features as four targets. The four feature types in the Discrete Wavelet Transform (DWT) step include: approximate sub-image (LL), vertical edge (HL), horizontal edge (LH), and diagonal feature (HH).

[0021] Furthermore, step (iv) specifically includes the following:

[0022] Step (four) Convert the YUV space back to RGB space to obtain the final processed image. The conversion is performed using the following formula:

[0023] R = Y + 0.956 * U + 0.620 * V,

[0024] G = Y - 0.272 * U - 0.647 * V,

[0025] B = Y - 1.108 * U - 1.705 * V.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] (1) This invention designs a frequency domain multi-target information steganography technique based on discrete wavelet transform. By using color space conversion and frequency domain space conversion, information triggers are designed for the frequency domain to achieve ideal concealment.

[0028] (2) This invention creatively utilizes the feature extraction capability of discrete wavelet transform to extend the information steganography target from a single target to a multi-target information steganography target, effectively improving the effectiveness of information steganography.

[0029] (3) This invention performs well in terms of information steganography effectiveness and concealment. At the same time, it has ideal evasion capabilities against a variety of advanced detection methods such as NeuralCleanse, STRIP and Fine-Pruning, and has ideal robustness. Attached Figure Description

[0030] Figure 1 This is a flowchart of a frequency domain multi-target information steganography technique based on discrete wavelet transform.

[0031] Figure 2 This is a diagram comparing the visual effects of the original image and the final processed image.

[0032] Figure 3 This is a schematic diagram of the STRIP test results.

[0033] Figure 4 This is a schematic diagram of the Fine-Pruning test results. Detailed Implementation

[0034] The present invention will now be described in detail through exemplary embodiments. However, it should be understood that, without further description, elements, structures, and features in one embodiment may be advantageously incorporated into other embodiments.

[0035] A frequency domain multi-target information steganography technique based on discrete wavelet transform specifically includes the following steps:

[0036] (I) Transforming RGB color images to YUV space and extracting UV chromaticity channels: RGB is the most commonly used color space in neural network models. The human eye perceives all three channels of an image as identically. However, the YUV space divides a color image into a luminance component (Y channel) and a chromaticity component (U and V channels), and the human visual system is less sensitive to the latter. Therefore, we transform the RGB color image to YUV space and select the UV channels of the image.

[0037] (II) Transforming the UV channels of the image to the frequency domain using DCT and embedding information triggers in the high-frequency region: DCT transforms the UV channels of the image from the spatial domain to the frequency domain. Small perturbations in the frequency domain correspond to large regions in the spatial domain. DCT has the advantage of good energy concentration; after DCT transformation, most of the energy of the image is concentrated in the low-frequency components, in the upper left part of the frequency map, while the high-frequency components are in the lower right part. The human visual system is less sensitive to high frequencies than low frequencies; therefore, we choose to embed information triggers in the high-frequency region by modifying the amplitude.

[0038] (III) Utilizing the feature extraction capabilities of Discrete Wavelet Transform (DWT), four types of features are used as four targets: Discrete Wavelet Transform (DWT) is an approximate wavelet transform method that decomposes a signal into approximate coefficients and detail coefficients at different scales. DWT uses discrete wavelet functions and discrete time scales, and achieves signal decomposition through filtering and downsampling operations. Typically, DWT decomposes a signal into a set of high-frequency and low-frequency sub-signals. The first-level decomposition formula of DWT is as follows:

[0039]

[0040] Where A1(n) are low-frequency (approximate) coefficients, and D1(n) are high-frequency (detail) coefficients. h(k) and g(k) are the low-pass and high-pass filter coefficients of the wavelet analysis filter, respectively. The DWT transform allows the decomposition and reconstruction of the signal at different scales to analyze different frequency components.

[0041] Performing a DWT transform on an image is essentially equivalent to repeatedly filtering different information from the image using low-pass and high-pass filters. After the DWT transform, the image yields four types of features: approximation subimage (LL), vertical edges (HL), horizontal edges (LH), and diagonal features (HH). By concatenating a sub-band of one of the four DWT features from the pre-processed image with the other three DWT feature sub-bands from the original image, all four objectives are successfully achieved.

[0042] (iv) Convert the YUV space back to the RGB space to obtain the final processed image, and then train it in the neural network model to evaluate the effectiveness and concealment of the information steganography method: Finally, perform a discrete wavelet inverse transform (IDWT). Since the neural network model is mainly trained in the RGB color space, the YUV channel image is converted back to the RGB channel to obtain the final processed image and train it in the neural network model.

[0043] To further illustrate the advantages of the above-described method of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0044] Example:

[0045] See Figure 1 , Figure 1 This is a flowchart of a frequency domain multi-target information steganography technique based on discrete wavelet transform, which illustrates the four steps of the above method in detail. Figure 2 The diagram shows a visual comparison between the original image and the final processed image. As can be seen from the diagram, the final processed image has no obvious noise points and is very close to the original image. Therefore, our method achieves a good visual concealment effect. Figure 3 The figure shows the detection results of applying the STRIP detection method to the method. As can be seen from the figure, the entropy distribution of the model prediction results is very similar when processing the original sample and the processed sample, which successfully avoids the detection of the STRIP detection method, indicating that the method maintains good concealment under STRIP detection. Figure 4 The figure shows the detection results of applying the Fine-Pruning detection method. As can be seen from the figure, with the increase of the number of branches, the decrease in embedding success rate (ESR) is less than the decrease in benign accuracy (BA). This indicates that the method has good effectiveness. It retains its functionality in the face of Fine-Pruning detection and is not easily detected or eliminated.

[0046] As shown in Table 1, PSNR measures the degree of image distortion; a higher value indicates less image distortion. SSIM assesses image structural similarity; a value closer to 1 indicates higher structural similarity between images. LPIPS learns perceptual image patch similarity; a value closer to 0 indicates that the image is closer to the original image. The data in the table shows that our method achieves ideal results in all three metrics and demonstrates excellent concealment.

[0047] Table 2 shows the effectiveness evaluation results on the Animal10 dataset. Under different neural network models, the ESR of all four targets is close to 100%, and the BA does not decrease significantly, indicating that the method has achieved ideal effectiveness.

[0048] Table 3 shows the effectiveness evaluation results on the Cifar10 dataset. Under different neural network models, the ESR of all four objectives is close to 100%, and the BA does not decrease significantly, indicating that the method has achieved ideal effectiveness.

[0049] Table 4 shows the effectiveness evaluation results on the Cifar100 dataset. Under different neural network models, the ESR of all four objectives is close to 100%, and the BA does not decrease significantly, indicating that the method has achieved ideal effectiveness.

[0050] As shown in Table 5, the average ESR, BA, and the evaluation results of the three stealth metrics (PSNR, SSIM, and LPIPS) obtained by averaging the ESR of the four targets on three datasets using our processing method are compared with those of several existing information steganography methods. The results show that our method achieves better results. In particular, the three stealth metrics achieve significantly superior results.

[0051] As shown in Table 6, the Neural Cleanse detection method uses an anomaly index to measure the deviation of the trigger size. If the model's anomaly index is greater than 2, detection is successful. The results in the table show that for all three datasets, the anomaly index of our information steganography method is less than the threshold of 2, indicating that it can circumvent the Neural Cleanse detection method.

[0052] As shown in Table 7, NAD uses a teacher network trained on a small subset of clean data to guide the fine-tuning of the background student network, thus eliminating the influence of triggers. The results in the table show that our method's ESR remains high after applying NAD, indicating that NAD is ineffective for our approach.

[0053] Table 1. Concealment Assessment Results under Three Indicators

[0054]

[0055] Table 2. Results of the validity evaluation of the Animal10 dataset.

[0056]

[0057] Table 3. Results of Cifar10 dataset validity evaluation

[0058]

[0059] Table 4. Validity Evaluation Results of the Cifar100 Dataset

[0060]

[0061] Table 5 shows the comparison results with existing processing methods.

[0062]

[0063] Table 6 Neural Cleanse Detection Results

[0064]

[0065] Table 7. Results of NAD treatment

[0066]

[0067] The embodiments described above are merely illustrative of the present invention and are not intended to limit the scope of protection of the present invention. Any simple modifications and variations made by those skilled in the art within the scope of the technical solutions described in the present invention should be included within the scope of the above-mentioned patent applications.

Claims

1. A frequency domain multi-target information steganography technique based on discrete wavelet transform, characterized in that, This method considers the concealment of information triggers in the frequency domain, sets multiple targets, and achieves extremely high concealment while maintaining ideal information steganography effectiveness. The multi-target information steganography technique includes the following steps: (a) Convert the RGB color image to YUV space and extract the UV chromaticity channels; (ii) Transform the UV channels of the image to the frequency domain using DCT and embed information triggers in the high-frequency region; (iii) Utilizing the feature extraction capability of discrete wavelet transform, the four types of features are used as four targets; (iv) Convert the YUV space back to the RGB space to obtain the final processed image, and then train it in a neural network model to evaluate the effectiveness and concealment of the information steganography method.

2. The frequency domain multi-target information steganography technique based on discrete wavelet transform according to claim 1, characterized in that, In step (one), the conversion from RGB color image to YUV color space is performed using the following formula: Y = 0.299 * R + 0.587 * G + 0.114 * B, U = 0.596 * R - 0.272 * G - 0.321 * B, V=0.212*R-0.523*G-0.311*B.

3. The frequency domain multi-target information steganography technique based on discrete wavelet transform according to claim 1, characterized in that, In step (ii), the UV channel of the image is transformed to the frequency domain by DCT, and the trigger is embedded in the high-frequency region. This is based on the characteristic that the human visual system is less sensitive to high-frequency components than to low-frequency components.

4. The frequency domain multi-target information steganography technique based on discrete wavelet transform according to claim 1, characterized in that, In step (iii), the feature extraction capability of discrete wavelet transform is used to select four types of features as four targets. The four feature types in the discrete wavelet transform (DWT) step include: approximate sub-image (LL), vertical edge (HL), horizontal edge (LH), and diagonal feature (HH).

5. The frequency domain multi-target information steganography technique based on discrete wavelet transform according to claim 1, characterized in that, In step (four), the YUV space is converted back to RGB space to obtain the final processed image. The conversion is performed using the following formula: R = Y + 0.956 * U + 0.620 * V, G = Y - 0.272 * U - 0.647 * V, B = Y - 1.108 * U - 1.705 * V.

6. The frequency domain multi-target information steganography technique based on discrete wavelet transform according to claim 1, characterized in that, In terms of information steganography effectiveness, it can achieve a high embedding success rate (ESR) of steganographic information, and the benign accuracy (BA) does not decrease significantly when predicting unprocessed inputs. In terms of information steganalysis concealment, the processed image is visually highly similar to the original image and performs well in metrics such as peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and learned perceptual patch similarity (LPIPS). In terms of robustness, it can effectively circumvent a variety of advanced information steganalysis detection methods.

Images