An Iterative Two-Stage Probability Adjustment Strategy With Progressive Incremental Searching for Image Steganography

Fan Wang, Xiang Zhang, Zhangjie Fu

Abstract: Adversarial example-based steganographic methods that utilize the gradients of target steganalyzer to update symmetric costs are emerging. The existing adversarial adjustment strategies for costs still have limited improvements in steganographic security. The existing gradient selection scheme, which sets a fixed gradient selection ratio for all images, is not delicate enough. To address the above problems, this paper proposes an iterative two-stage probability adjustment strategy with a progressive incremental searching mechanism (ITPA-PIS) to further improve the security of updated asymmetric distortions. Unlike previous works that adopted the cost as the adjustment object, we explore a new adjustment object, i.e., probability, and then design an iterative two-stage probability adjustment strategy (ITPA) to obtain a more secure asymmetric distortion, thereby improving the anti-detection performance of the traditional symmetric distortion algorithms against deep learning-based steganalyzers. In addition, we specifically design a progressive incremental searching mechanism (PIS) to select partially efficient gradients to guide the probability adjustment. Unlike existing gradient selection schemes that manually set a fixed selection ratio, PIS adopts a progressive searching method to dynamically determine the gradient selection ratio suitable for each image, thereby enhancing the overall performance of the proposed ITPA again. The experimental results show that our proposed ITPA-PIS achieves outstanding security performance on the CNN-based steganalysis models XuNet, YedroujNet, SRNet, and EfficientNet and hand-crafted feature-based steganalysis models SRM and MaxSRMd2 under the adversary unawareness and adversary awareness scenarios.

External IDs:doi:10.1109/tcsvt.2024.3398222