Abstract:Current deep learning-based steganalyzers often depend on specific image dimensions, leading to inevitable adjustments in network structure when dealing with varied image sizes. This impedes their effectiveness in managing the wide range of image sizes commonly found on social media. To address this issue, our paper presents a novel steganalytic network that is optimized for fixed-size (notably, ) color images, but is capable of efficiently detecting stego images of arbitrary size without needing retraining or modifications to the network. Our proposed network is comprised of four modules. In the initial stem module, we calculate truncated residuals for each color channel of the input image. Diverging from existing steganalytic networks that rely on vanilla convolution, we have developed a pixel difference convolution module designed to better capture the artifacts introduced by steganography. Following this, we introduce an enhanced Transformer module with selective pooling, aimed at more effectively extracting global steganalytic features. To guarantee our network's adaptability to different image sizes, we have developed a selective pooling strategy. This involves using global covariance pooling for fixed-size color images and spatial pyramid pooling for color images of various other sizes. This approach effectively standardizes the feature maps into uniform feature vectors. The final module is focused on classification. Extensive testing results on the ALASKA II color image dataset have demonstrated that our approach significantly improves detection performance for both fixed-size and arbitrary-size images, achieving state-of-the-art results. Additionally, we provide numerous ablation studies to confirm the effectiveness and soundness of our proposed network architecture.

Image Steganalysis Network Based on Dual-Attention Mechanism

A Dct-Based Image Steganographic Method Resisting Statistical Attacks

Highly Accurate End-to-end Image Steganalysis Based on Auxiliary Information and Attention Mechanism

A Blind Steganalytic Scheme Based on DCT and Spatial Domain for JPEG Images.

Iscmis:Spatial-Channel Attention Based Deep Invertible Network for Multi-Image Steganography

Channel Attention Image Steganography With Generative Adversarial Networks

StegaVision: Enhancing Steganography with Attention Mechanism

Spatial Steganalysis Based on Non-Local Block and Multi-Channel Convolutional Networks.

Embedding Probability Guided Network for Image Steganalysis.

Generative adversarial networks-based image steganography with multiscale features integration

Deeply‐Recursive Attention Network for Video Steganography

Depth-Wise Separable Convolutions and Multi-Level Pooling for an Efficient Spatial CNN-Based Steganalysis.

A Novel Technique for Image Steganalysis Based on Separable Convolution and Adversarial Mechanism

A Siamese CNN for Image Steganalysis

Maximizing steganalysis performance using siamese networks for image

A High-Capacity Steganography Algorithm Based on Adaptive Frequency Channel Attention Networks

Preemptive Image Protection against Steganography

CNN-Assisted Steganography -- Integrating Machine Learning with Established Steganographic Techniques

Text Steganalysis with Attentional LSTM-CNN

Color Image Steganalysis Based on Pixel Difference Convolution and Enhanced Transformer With Selective Pooling

Text Steganalysis Based on Hierarchical Supervised Learning and Dual Attention Mechanism.