Enhanced image steganalysis through reinforcement learning, generative adversarial networks, and vision transformers (RGV-Stega)

As steganographic techniques advance and object data are increasingly generated in various domains, detecting concealed data embedded in digital media has become more challenging; thus, image steganalysis plays an important role. Conventional detection techniques are becoming less effective against these evolving strategies, which calls for better solutions. In this paper, we present RGV-Stega, a new high-level structure that combines Reinforcement Learning (RL), Generative Adversarial Networks (GANs), and Vision Transformers (ViTs) to further enhance the detection of steganographic contents in images. Artificial intelligence techniques such as RL, GANs, and ViTs play significant roles in current steganalysis work: RL helps an agent learn optimal feature extraction strategies, GANs yield rich steganographic samples for strong training, and ViTs help capture long-range dependencies in image data, thereby boosting top-1 accuracy. On benchmark datasets such as BOSSBase and BOWS, our method RGV-Stega achieves up to 93% accuracy, comparable to conventional CNN- and ViT-based models. Findings show that combining RL, GANs, and ViTs is an effective approach to addressing issues arising from advances in steganographic techniques. This improves steganalysis’s ability to identify hidden content.