A Steganalysis Method Based on Mobilevit-v3

Abstract

The present invention provides a steganalysis method based on Mobilevit-v3, comprising the following steps: making a dataset - preprocessing - feature extraction - feature classification. The present invention is based on the Mobilevit-v3 Block having better generalization ability and better robustness; based on the CFP Block having fewer parameters and better performance.

Description

Technical Field

[0001] The present invention relates to a steganalysis method based on Mobilevit-v3, belonging to the field of content security technology in cyberspace security. Background Art

[0002] Image steganalysis technology and image steganography are in a reciprocal relationship. Image steganalysis technology classifies whether secret information has been embedded in the pictures spread in public media by image steganography, and these information are indistinguishable to the naked eye.

[0003] The Mobilevit-v3 Block makes simple and effective improvements to the fusion block. First, replace the 3×3 convolutional layer with a 1×1 convolutional layer; second, fuse the features of the local representation block and the global representation block together, rather than fusing the input and the global representation block together; third, add the input feature as the last step in the fusion block before generating the output of the Mobilevit Block; finally, replace the ordinary 3×3 convolutional layer with a depthwise 3×3 convolutional layer in the local representation block, solving the scaling problem and simplifying the learning task.

[0004] The CFP Block combines the Inception module and the dilated convolution module, and is called the Channel-wise Feature Pyramid (CFP) module. This module jointly extracts feature maps and context information of various sizes, significantly reducing the number of parameters.

[0005] Existing adaptive image steganography algorithms can hide more secret information in the complex texture areas of images and make the statistical characteristics of the modified images close to those of the original images, but they will reduce the network performance, use a large number of parameters, resulting in a low detection accuracy of the model. Summary of the Invention

[0006] To solve the above technical problems, the present invention provides a steganalysis method based on Mobilevit-v3, which can improve the network performance and reduce the number of parameters.

[0007] The present invention is achieved through the following technical solutions.

[0008] A steganalysis method based on Mobilevit-v3 provided by the present invention includes the following steps:

[0009] ① Make a dataset of stego images;

[0010] ② Preprocess the images in step ①;

[0011] ③After step ②, the high-frequency information of the image is retained, and the Mobilevit-v3 Block and CFP Block are used to extract features from the image obtained after preprocessing in step ②;

[0012] ④After feature extraction, a binary classification is performed on whether the image is a stego image according to the probability value.

[0013] In the said step ①, three adaptive steganography algorithms, WOW, S-UNIWARD, and HILL, and four payloads of payload = 0.1bpp, 0.2bpp, 0.3bpp, and 0.4bpp are used to perform steganography on the BOSSbase v1.01 original dataset to obtain multiple different stego image datasets.

[0014] In the said step ②, the SRM high-pass filter is used for image preprocessing to obtain a noise residual image.

[0015] In the said step ④, a fully connected layer using Softmax is used to classify the extracted features.

[0016] The beneficial effects of the present invention are as follows: Based on the Mobilevit-v3 Block, it has better generalization ability and better robustness; based on the CFP Block, it has fewer parameters and better performance. Description of the Drawings

[0017] Figure 1 It is a network model diagram designed based on the Mobilevit-v3 Block and CFP Block of the network structure of the present invention;

[0018] Figure 2 It is the network structure diagram of the Mobilevit-v3 Block of the present invention;

[0019] Figure 3 It is that the CFP Block of the present invention consists of K feature maps with different dilation rates. Detailed Embodiments

[0020] The technical solutions of the present invention are further described below, but the scope of protection claimed is not limited thereto.

[0021] A steganalysis method based on Mobilevit-v3 includes the following steps:

[0022] ①Make a dataset of stego images;

[0023] ②Preprocess the images in step ①;

[0024] ③After step ②, the high-frequency information of the image is retained, and the Mobilevit-v3 Block and CFP Block are used to extract features from the image obtained after preprocessing in step ②;

[0025] ④After feature extraction, the image is binary-classified according to the probability value.

[0026] In step ①, three adaptive steganography algorithms, WOW, S-UNIWARD, and HILL, and four payloads of 0.1 bpp, 0.2 bpp, 0.3 bpp, and 0.4 bpp are used to perform steganography on the BOSSbase v1.01 original dataset to obtain multiple different stego-image datasets.

[0027] Preferably, 10 or 12 different stego-image stego datasets are obtained.

[0028] In step ②, the SRM high-pass filter is used for image preprocessing to obtain the noise residual image.

[0029] Preferably, the residual image is obtained through 30 SRM high-pass filters.

[0030] ④After feature extraction, it is binary-classified according to the probability value whether the image is a stego image.

[0031] Further, as Figure 1 shown, is the network model of the Mobilevit-v3 Block and CFP Block. As Figure 2 shown, in the present application, a 1×1 convolutional layer is used in the fusion block of the Mobilevit-v3 Block. By changing the network width and keeping the depth constant, a large increase in parameters and FLOP during scaling can also be avoided. In the fusion layer, the features from the local and global representation blocks are concatenated in the Mobilevit-v3 Block to fuse the local and global features, and the input features are added to the output of the 1×1 convolutional layer in the fusion module. To further reduce the parameters, the 3×3 convolutional layer of the local representation block is replaced with a depthwise 3×3 convolutional layer. As Figure 3 shown, the CFP Block consists of K feature map channels with different dilation rates. A 1×1 convolution is used for the input (mapping from high dimension to low dimension) to reduce the dimension from M to M / K; then there are the first to the third dimension-asymmetric blocks of M / 4K, M / 4K, and M / 2K. Starting from the second channel, the feature maps are gradually combined using the summation operation, and then they are concatenated to construct the final hierarchical feature map. The input size of this network structure is 256*256*1, and the output is the probabilities of non-stego images and stego images.

[0032] In summary, the present invention uses Mobilevit-v3 Block and CFP Block to improve network performance and reduce the number of parameters, enabling the model detection to achieve better accuracy.

Claims

1. A steganalysis method based on Mobilevit-v3, characterized in that: It includes the following steps: ① Making a dataset of stego images; ② Preprocessing the images in step ①; ③ After step ②, the high-frequency information of the images is retained, and the Mobilevit-v3 Block and CFP Block are used to extract features from the images obtained after preprocessing in step ②; ④ After feature extraction, a binary classification is performed on whether the image is a stego image according to the probability value; The CFP Block consists of K feature map channels with different dilation rates, and the input is reduced from M dimensions to M / K dimensions using a 1×1 convolution; Then there are the first to the third dimension-asymmetric blocks M / 4K, M / 4K, and M / 2K. Starting from the second channel, the feature maps are gradually combined using the summation operation, and then they are connected to construct the final hierarchical feature map, and the output is the probability of non-stego images and stego images.

2. The steganalysis method based on Mobilevit-v3 according to claim 1, characterized in that: In step ①, three adaptive steganography algorithms, WOW, S-UNIWARD, and HILL, and four payloads payload = 0.1bpp, 0.2bpp, 0.3bpp, and 0.4bpp are used to perform steganography on the BOSSbase v1.01 original dataset to obtain multiple different stego image datasets.

3. The steganalysis method based on Mobilevit-v3 according to claim 1, characterized in that: In step ②, the SRM high-pass filter is used for image preprocessing to obtain the noise residual image.

4. The steganalysis method based on Mobilevit-v3 according to claim 1, characterized in that: In step ④, a fully connected layer using Softmax is used to classify the extracted features.

Images