AdvSGAN: Adversarial image Steganography with adversarial networks

摘要

Steganalysers based on deep learning achieve state-of-the-art performance. However, due to the difficulty of capturing the distribution of the high-dimensional covers, traditional steganography schemes construct more complex artificial rules within expert knowledge, which is usually challenging to obtain to counter these powerful steganalysers. Adversarial learning is a valuable potential for steganography. There have some steganography schemes through playing an adversarial game within deep neural networks. However, there is a vast security margin needed to reduce. In this paper, we propose AdvSGAN, which learns an image steganography scheme represented by a restricted neural coder from scratch by playing an adversarial game between the restricted neural coder and adversaries in the adversary model, i.e., In-Training and Out-Training adversaries. The restricted neural coder is implemented by two neural networks named SE and SD are to perform encoding and decoding transformation respectively, and a flexible restriction model to constrain the covers' embedding space to improve the performance. The In-Training adversary is implemented by another network of discriminator named Eve in the GANs model. The Out-Training adversary is implemented by the targeted CNN based steganalyser. By playing adversarial game jointly with Eve, SE and SD are evolving to find the possible transformation. Meanwhile, by attacking the Out-Training adversary in a white-box setting, the obtained gradient provides instructive guidance for evolving to find the optimal steganographic scheme. Experiments demonstrate that the proposed steganographic scheme achieves better security performance even in high capacity against targeted steganalyser, and still has some transferability to other unaware steganalysers.