Reversible adversarial steganography for security enhancement

Abstract

In adaptive steganography, existing universal stego post-processing methods can enhance security, but suffer from incorrect extraction of messages and undesired visual defects, especially in flat regions. To address this issue, a reversible adversarial steganography method is proposed by modifying the LSB or 2ndLSB of stego-images, which has promising visual quality and security. To that end, the content-adaptive adversarial perturbations are first generated, which consider image texture with noise residual features to counter deep learning-based steganalyzers. Then, a data compression strategy of adversarial perturbations is designed by applying lossless run-length encoding based on the sparse nature of non-zero elements in the perturbations to reduce the perturbation’s quantity. Finally, reversible data hiding based on ternary coding is applied to embed and extract stego images with compressed adversarial perturbations. Extensive experimental results demonstrate that the proposed method can effectively enhance security and visual quality compared with state-of-the-art methods.