Robustness enhancement against adversarial steganography via steganalyzer outputs

Recently, CNN (convolutional neural network) steganalyzers have significantly outperformed handcrafted features in detecting steganography. However, adversarial steganography has challenged the applications of them in the real world. Adversarial steganography can easily deceive the target CNN steganalyzer while sending secret messages. In this paper, a general framework is proposed. It that can improve the robustness of CNN steganalyzers against adversarial steganography while keeping detecting cover and conventional stego images accurately. Specifically, a rough filter that filters adversarial stego images out of the input data is set. It exploits the differences between cover and adversarial stego images on probabilistic outputs of the target CNN steganalyzer and a handcrafted steganalyzer. Extensive experiments show that the proposed framework can significantly improve the robustness of CNN steganalyzers. In the real-world scenario where cover, conventional stego and adversarial stego images are mixed, the robustness enhanced CNN steganalyzers can achieve the optimal overall performance.