Active Defense Against Image Steganography

Image steganalysis is the art and science of detecting whether an image contains secrete messages or not, which can prevent malicious usage of steganography. However, steganalysis belongs to passive defense, in the sense that it can only be applied after the stego image is generated. Therefore, there still exists loophole that secrete messages communication could already been accomplished when the stego image is detected. To eliminate malicious steganography from the source, in this paper, an active defensive framework for deep image steganography called ADPI (Active Defense based on Perturbation Injection) is proposed, wherein a defender competes against a steganographer to learn the active defensive strategy. Specifically, on the side of the defender, a generator is adopted to take the original cover image as input, and learn the imperceptible perturbation map. Such perturbation map is added with the original cover image as the enhanced cover image. On the side of the steganographer, a steganographic network is applied to perform message embedding and extraction on the enhanced cover image. The key of ADPI is that the perturbation map is optimized with the goal of reducing the accuracy of the message recovery while maintaining its invisibility. By this means, the active defense can be launched in an effective and imperceptible manner. Experimental results show that the proposed ADPI can be applied to defend against various steganographic methods.