Deep learning for steganalysis: evaluating model robustness against image transformations.

This study investigates the robustness of deep learning-based steganalysis models against common image transformations because most literature has not paid enough attention to resilience assessment. Current and future applications of steganalysis to guarantee digital security are gaining importance regarding real-world modifications: resizing, compression, cropping, and adding noise. These included the following five basic models: EfficientNet, SRNet, ResNet, Xu-Net, and Yedroudj-Net. We evaluated these models' pre- and post-transformation performances based on various metrics like accuracy, precision, recall, F1-score, and AUC with the BOSSBase dataset. Our results showed that EfficientNet is the most robust among the considered architecture transformations. Still, it also underlined significant degradations for state-of-the-art models, Xu-Net and Yedroudj-Net, especially with added noise. These results indicate the need to develop more robust architectures capable of sustaining real-world image alterations. In practice, it will assist practitioners in choosing models that best suit operational environments and lay the necessary platform for future enhancements in the design of such models. In this regard, in the future, more transformations should be researched with ensemble and adaptive approaches to improve robustness further.