CN-121982360-A - Steganalysis method and device integrating multi-scale and frequency domain self-adaptive characteristics

CN121982360ACN 121982360 ACN121982360 ACN 121982360ACN-121982360-A

Abstract

The invention discloses a steganalysis method and a steganalysis device integrating multi-scale and frequency domain self-adaptive characteristics, the method firstly constructs and trains a steganalysis network, and then inputting the images to be detected into a network for classification, and judging whether the images are steganographic images or not by calculating probability values and comparing the probability values with threshold values. The steganalysis network comprises three core modules of preprocessing, feature extraction and classification. The preprocessing module realizes end-to-end characteristic enhancement through the multi-scale texture fusion module and the self-adaptive frequency attention mechanism. The multi-scale texture fusion module adopts multi-scale central difference convolution to adaptively extract noise residual errors, and the self-adaptive frequency attention mechanism filters low-frequency interference through frequency domain transformation and a leachable threshold. The feature extraction module captures and fuses the intensity features and the gradient features through the double-branch structure of the gradient-intensity fusion module. The classification module outputs a probability value based on the fusion feature. The invention has obviously improved steganalysis precision, efficiency and generalization capability.

Inventors

LIN HONGRUI
WENG SHAOWEI
ZHANG TIANCONG

Assignees

福建理工大学

Dates

Publication Date: 20260505
Application Date: 20251210

Claims (9)

1. A steganalysis method integrating multiscale and frequency domain self-adaptive features is characterized by comprising the following steps: constructing a steganography analysis network integrating multi-scale and frequency domain self-adaptive characteristics and training; Inputting the image to be detected into a trained steganography analysis network which integrates multi-scale and frequency domain self-adaptive characteristics to be classified, and calculating a probability value of the image to be detected containing steganography information; Judging whether the image to be detected is a steganographic image according to a comparison result of the probability value and a preset threshold value, and completing image steganographic analysis; The steganography analysis network for fusing the multi-scale and frequency domain self-adaptive features comprises a preprocessing module, a feature extraction module and a classification module, wherein the preprocessing module comprises a multi-scale texture fusion module and a self-adaptive frequency attention mechanism and is used for adaptively extracting noise residual errors from an input image and enhancing the signal to noise ratio of the steganography features to obtain a feature map, the feature extraction module comprises at least one gradient-intensity fusion module, the gradient-intensity fusion module adopts two branches to capture local intensity features and central differential gradient features of the feature map respectively and fuse the local intensity features and the central differential gradient features with the feature map after splicing to obtain fusion features, and the classification module is used for calculating and outputting probability values containing steganography information based on the fusion features.
2. The steganalysis method for fusing multi-scale and frequency domain self-adaptive features of claim 1, wherein the multi-scale texture fusion module comprises more than three parallel paths, each path is provided with a sequence consisting of a plurality of central differential convolutions with fixed convolution kernels, the central differential convolutions of different paths are provided with fixed convolution kernels with different sizes, and the outputs of all paths of the multi-scale texture fusion module are fused through batch normalization and channel dimension splicing operation.
3. The steganalysis method for fusing multi-scale and frequency domain self-adaptive features of claim 1, wherein three paths of the multi-scale texture fusion module are adopted, the convolution kernel size of the central differential convolution of each path is respectively fixed to be 3×3, 5×5 and 7×7, and the number of the central differential convolutions CDCs of each path is 10.
4. The steganalysis method combining multiscale and frequency domain adaptive features of claim 1, wherein the adaptive frequency attention mechanism comprises: the fast Fourier transform unit is used for converting the output of the multi-scale texture fusion module into frequency domain characteristics; an adaptive component selection unit that filters out low power components in the frequency domain features based on a leachable threshold θ; the inverse fast Fourier transform unit is used for converting the filtered frequency domain characteristics into a return space domain; a weight generation unit that generates spatial and channel attention weights by a3×3 convolution block, a point-by-point convolution, and a Sigmoid function; and the feature weighting unit is used for multiplying the original feature with the corresponding weight to obtain the enhanced feature.
5. The steganalysis method for fusing multiscale and frequency domain adaptive features of claim 1, wherein the gradient-intensity fusion module comprises: the upper branch is used for extracting local intensity features by using two 3 multiplied by 3 convolution blocks after the dimension is reduced by point-to-point convolution; The lower branch is subjected to point-by-point convolution for dimension reduction, and then a central differential convolution block is used for extracting central differential gradient characteristics, wherein lambda=1 of the central differential convolution block; and the fusion unit is used for fusing the output characteristics of the upper branch and the lower branch with the input characteristic graphs to obtain fusion characteristics after the output characteristics of the upper branch and the lower branch are spliced in the channel dimension.
6. The method for steganalysis of fusion of multi-scale and frequency domain adaptive features according to claim 1, wherein the classification module comprises a global pooling layer and a full connection layer, the fusion features are converted into discrimination vectors through the global pooling layer and the full connection layer, probability values of the steganalysis images are calculated on the basis of the discrimination vectors, and the steganalysis images are determined when the probability values are larger than a preset threshold.
7. The steganalysis method combining multi-scale and frequency domain adaptive features according to claim 1 or 6, wherein the preset threshold is 0.5.
8. The steganalysis method combining multi-scale and frequency domain adaptive features according to claim 1, wherein the steganalysis network combining multi-scale and frequency domain adaptive features is trained by a random gradient descent optimizer, momentum is set to 0.9, weight attenuation is 5×10 -4 , initial learning rate is 0.01, and attenuation is carried out in stages in the training process.
9. A steganalysis device for fusing multi-scale and frequency-domain adaptive features, comprising a computer device, the computer device comprising a processor and a memory, the memory storing computer instructions, the processor being configured to execute the computer instructions stored in the memory, the device implementing the steps implemented by the steganalysis method for fusing multi-scale and frequency-domain adaptive features of any one of claims 1 to 8 when the computer instructions are executed by the processor.

Description

Steganalysis method and device integrating multi-scale and frequency domain self-adaptive characteristics Technical Field The invention relates to the technical field of image processing, in particular to a steganalysis method and device for fusing multi-scale and frequency domain self-adaptive characteristics. Background Image steganography is a technique in which secret data is embedded by slightly modifying the spatial pixel values or frequency coefficients of a carrier image, the core goal of which is to achieve covert communication without drawing attention from unauthorized parties. In contrast, image steganalysis is aimed at mining statistical differences between normal and steganographic images, thereby detecting whether hidden information is present in suspicious images. Existing steganalysis methods are mainly divided into two types, namely a traditional method based on manual characteristics or a characteristic learning method based on deep learning. Early methods relied on manually designed statistical features such as statistical matrices, discrete cosine transform features, wavelet coefficient features, co-occurrence matrix features, and the like. These methods identify steganographic traces by combining classifiers, but the feature extraction process is highly dependent on expert experience and difficult to capture complex nonlinear statistical properties. Conventional manual features or single convolution operations have difficulty in comprehensively capturing subtle statistical anomalies introduced by steganography, and particularly have limited signal-to-noise ratio (SNR) improvement under low frequency component interference. With the development of deep learning, convolutional Neural Networks (CNNs) are widely used for steganalysis. Although the precision of a complex network structure (such as a deep residual error module) can be improved, the training and reasoning process is time-consuming and energy-consuming, and is difficult to apply to a resource-limited scene. Disclosure of Invention The invention aims to provide a steganography analysis method (SAFNet) and a steganography analysis device which are integrated with multi-scale and frequency domain self-adaptive characteristics, wherein steganography analysis is completed by combining a learnable multi-scale convolution and frequency domain selection mechanism from a spatial domain image. The technical scheme adopted by the invention is as follows: a steganalysis method integrating multiscale and frequency domain self-adaptive features comprises the following steps: constructing a steganography analysis network integrating multi-scale and frequency domain self-adaptive characteristics and training; Inputting the image to be detected into a trained steganography analysis network which integrates multi-scale and frequency domain self-adaptive characteristics to be classified, and calculating a probability value of the image to be detected containing steganography information; Judging whether the image to be detected is a steganographic image according to a comparison result of the probability value and a preset threshold value, and completing image steganographic analysis; The steganography analysis network for fusing the multi-scale and frequency domain self-adaptive features comprises a preprocessing module, a feature extraction module and a classification module, wherein the preprocessing module comprises a multi-scale texture fusion module (MTFM) and a self-adaptive frequency attention mechanism (AFA) and is used for adaptively extracting noise residual errors from an input image and enhancing the signal to noise ratio of the steganography features to obtain a feature map, the feature extraction module comprises at least one gradient-intensity fusion module (GIFM), the gradient-intensity fusion module (GIFM) adopts two branches to capture local intensity features and central differential gradient features of the feature map respectively, and the local intensity features and the central differential gradient features are fused with the feature map after being spliced to obtain fusion features, and the classification module is used for calculating and outputting probability values containing steganography information based on the fusion features. Specifically, the feature extraction includes two residual blocks (RestNetBlock) and three gradient-intensity fusion modules. Further, the multi-scale texture fusion module (MTFM) comprises more than three parallel paths, each path is provided with a sequence consisting of a plurality of Central Differential Convolutions (CDCs) with fixed convolution kernels, the central differential convolutions of different paths are provided with the fixed convolution kernels with different sizes, and the outputs of all paths of the multi-scale texture fusion module (MTFM) are fused through Batch Normalization (BN) and channel dimension splicing operation. Further, the expression of the center different