Explainable Hybrid Gan–Rednet Framework For Bit-Level Medical Image Steganalysis

Abstract

With the increasing digitization of medical imaging in healthcare, protecting sensitive patient data has become critical. Steganography, the process of hiding information within images, poses significant security risks, necessitating advanced steganalysis techniques for medical images. Traditional methods rely on handcrafted features and shallow learning, limiting their performance on high-resolution and bit-level embedded data. This study proposes an Explainable Hybrid GAN–REDNet Framework for bit-level medical image steganalysis and security enhancement. The framework integrates generative adversarial networks (GAN) for encoding-decoding and data augmentation with Residual Enhanced Discriminative Networks (RED-Net) for precise stego detection, ensuring both high fidelity and robust classification. A multimodal brain tumor dataset (CT and MRI) was used, comprising 5,000 images. Preprocessing included normalization, augmentation, and generation of stego images via multi-bit LSB embedding. GAN-based augmentation enriched feature diversity, while RED-Net discriminated cover and stego images at the pixel and bit levels. Bit-plane sensitivity analysis identified lower bit-planes (0–2) as critical for detection. Explainable AI techniques, including GradCAM, were employed to interpret model focus areas. The hybrid framework achieved 86.11% accuracy, 0.86 F1-score, and ROC-AUC of 0.9989, outperforming baseline REDNet (78.42% accuracy). Confusion matrix analysis confirmed high stego detection sensitivity, and GradCAM visualizations aligned with bit-plane sensitivity findings, validating interpretability. The proposed model enhances medical image security by combining generative and discriminative paradigms, enabling robust, interpretable, and bit-level steganalysis, critical for ethical and secure healthcare imaging.