Generating JPEG Steganographic Adversarial Example via Segmented Adversarial Embedding

Abstract

Nowadays, Convolutional Neural Network (CNN) based steganalytic schemes further improves the detection ability of the steganalyzer comparing with feature based schemes. Besides steganalysis, CNN model can also be used in steganography. Inheriting the mechanism from adversarial attack to CNN model, adversarial embedding is a kind of steganographic scheme that exploits the knowledge of CNN-based steganalyzer. Adversarial embedding can effectively improve security performance of typical adaptive steganographic schemes. In this paper, we propose a novel adversarial embedding scheme for steganography named as Segmented Adversarial Embedding (SAE). The core of SAE is separating the embedding process into several partial embedding processes and performing adversarial embedding in each segment. In each partial embedding process, there is an individual CNN model corresponding to the current embedding stage. In the embedding process, a novel scheme is applied in the cost adjustment. Comparing with the adjustment scheme that utilizes the gradient sign, the new scheme also takes the gradient magnitude into account, which further makes use of the gradient information. Besides the typical implementation of SAE, we also develop a simplified variant with lower complexity. The evaluations on different kinds of steganalyzer prove that SAE is effective to improve the performance of existing steganographic scheme.

X. Zhao—This work is supported by the State Grid Corporation of China Project (5 700-202018268A-0-0-00).