Improving the robustness of steganalysis in the adversarial environment with Generative Adversarial Network

As technology advances, the steganalysis methods based on Convolutional Neural Network (CNN)can more precisely detect steganography behavior. However, the existing CNN-based steganalysis methods have the problem of insufficient robustness in adversarial environments. Specifically, the steganography methods take advantage of adversarial examples' capacity to deceive deep learning models, adding small perturbations to the stego images to generate adversarial stego images that can successfully evade the detection of CNN-based steganalysis. In this work, we design a robust steganalysis based on the Generative Adversarial Network (GAN), namely RS-GAN. The proposed method reconstructs the suspicious images after learning the distribution of clean images to identify adversarial stego images. Specifically, RS-GAN first disentangles suspicious images to obtain the texture and structure vectors. Afterward, reconstructed images are synthesized based on the above vectors, which have similar distributions to the clean images. In addition, we design an anomaly score calculation module to obtain the reconstruction loss and the discriminator loss and then determine the anomaly scores. Finally, the RS-GAN compares the anomaly scores with the threshold to eliminate the anomaly examples. According to the experimental comparison and analysis, RS-GAN can effectively eliminate the adversarial stego images and increase the robustness of steganalysis in adversarial environments. In the LSUN dataset, when the Fast Gradient Method is employed as the attack with a perturbation value of 0.0013, the accuracy of YeNet decreased to 47.867%, while RS-GAN maintains an accuracy of 72.100%.