Adaptive region assisted GAN for image steganography

Jingwen Cai, Fen Xiao, Kehan Zhang, Xieping Gao

Abstract: Generative adversarial network is recognized as an effective tool for hiding the secret information inside an image in a visually imperceptible way. Previous studies still suffered from challenges such as cover selection, anomalous pixels disturbance, imbalance between capacity, security and accuracy. This paper proposes a GAN-based adaptive local image steganography framework that adaptively selects hidden regions for encoding and employs a dual context discriminator to optimize the network. Specifically, the Cross-Channel Hybrid Local Binary Pattern is introduced and then utilized to measure the texture complexity of each cover region. In the generation stage, the region with the highest complexity is selected as the optimal site for embedding the secret data. The generative network is optimized with the proposed dual contextual discriminator that considers local and global information to maintain the authenticity and global consistency of the stego image. A weighted Mean Squared Error (MSE) joint loss function is designed for training generative adversarial network. The experimental results show that the proposed method outperforms existing steganography techniques on datasets such as LSUN, CelebA, and Place-365 Standard, both in terms of security and capacity.

External IDs:dblp:journals/mms/CaiXZG25