CNN-based Steganalysis Detects Adversarial Steganography via Adversarial Training and Feature Squeezing
Peng Ye, Guobin Fu, YingGuang Luo, Qi Jun Yu, Lijuan Wang · 2023
Deep learning model's vulnerability and unreliability cannot be ignored. For example, adversarial steganography exploits adversarial examples to escape the recognition of CNN-based steganalyzer, which is a pressing issue for intelligent steganalysis algorithms. In this paper, we propose a steganalysis method that can effectively detect adversarial steganography via adversarial training and feature squeezing. Firstly, we train the steganalyzer based on the adversarial cover and stego images with small perturbations to improve the robustness of the steganalyzer, which can effectively identify small perturbations of samples. Additionally, the proposed steganalysis decreases the utilization space of adversarial attacks by squeezing the input samples' features throughout the detection phase. The results of our experiments demonstrate that the accuracy of detecting adversarial steganography may be effectively improved by the proposed approach.