Adversarial-robust steganalysis system leveraging adversarial training and EfficientNet

Abstract

Steganalysis aims to detect hidden messages within digital media, presenting a significant challenge in the field of information security. This paper introduces an adversarial-robust steganalysis system leveraging adversarial training and the powerful feature extraction capabilities of EfficientNet. We utilize EfficientNet to extract robust features from images, which are subsequently classified by a dense neural network to distinguish between steganographic and non-steganographic content. To enhance the system’s resilience against adversarial attacks, we implement a custom adversarial training loop that generates adversarial examples using the fast gradient sign method (FGSM) and integrates these examples into the training process. Our results demonstrate that the proposed system not only achieves high accuracy in detecting steganographic content but also maintains robustness against adversarial perturbations. This dual approach of leveraging state-of-the-art deep learning architectures and adversarial training provides a significant advancement in the field of steganalysis, ensuring more reliable detection of hidden messages in digital images.