SA ‐Net‐Stego‐Medical: Explainable and Adversarially Robust Medical Image Steganalysis Using Self‐Attention Networks

The regulation has required regulatory‐compliant steganalysis to securely deliver medical images in telemedicine with a high level of accuracy in detection, resistance to adversarial attacks, and clinical interpretability. Although recent LSTM‐CNN hybrids offer parameter efficiency, they do not provide pixel‐level explainability, global spatial modeling, and can be compared to adaptive attacks that are essential obstacles to the usage of FDA (Food and Drug Administration). We present SA‐Net‐Stego‐Medical a new four‐stage model that is trained on clinically realistic datasets: ChestMNIST chest x‐rays and PathMNIST pathology slides. SA‐Net (Self‐Attention Network) replaces sequential LSTM (Long‐Short Term Memory) with multi‐head self‐attention, adding CBAM (Convolutional Block Attention Module) attention to SRM (Spatial Rich Model) residuals, and using the Grad‐CAM++ explainability, outperforms baselines significantly in both datasets and also shows excellent cross‐domain generalization. Grad‐CAM++ heatmaps accurately identify steganographic features to embedding‐prone anatomic structures (lung parenchyma, gland tissue) and get high plausibility scores with board‐certified radiologists. Adversarial training is resistant to FGSM (Fast Gradient Sign Method) and PGD (Projected Gradient Descent) attacks, and is significantly more resistant than baseline methods. There is the first production‐ready platform of secure, explainable medical imaging security in adversarial telemedicine setting with compact parameterization, real‐time inference, and full FDA 21 CFR (Code of Federal Regulations) Part 11 transparency.