Deep reinforcement learning-based reversible medical image encryption framework for secure IoMT environments

The Internet of Medical Things (IoMT) environments face significant challenges in securely transmitting and storing medical images due to limited computational resources, multiple device types, and increasing cybersecurity threats. This paper describes a reversible RGB medical image encryption framework that employs deep reinforcement learning by combining adaptive policy learning with deterministic cryptographic algorithms. A Deep Q-Network (DQN) is used to dynamically select encryption actions based on statistical features extracted from the intermediate encrypted image state. To achieve strong security and precise image recovery, the framework employs a multi-layer reversible technique that comprises SHA-512-based keystream masking, Arnold scrambling with padding preservation, and chaotic diffusion. Extensive testing shows that this technique achieves high entropy, virtually optimum Number of Pixel Change Rate (NPCR) and Unified Average Changing Intensity (UACI) metrics, minimal pixel correlation and near-zero Structural Similarity Index Measure (SSIM) between the original and encrypted images, indicating a robust protection against statistical and differential attacks. Furthermore, the framework is robust against noise, data loss, occlusion, chosen plaintext, and determinism leaking attacks. Unlike fixed chaos-based encryption systems, the proposed framework introduces reinforcement learning-based adaptive action selection within a strictly reversible cryptographic pipeline. The effective key space exceeds 2⁵¹² due to SHA-512-based seed derivation and nonce-driven randomness. The overall computational complexity of the encryption process is O(H × W × T), making it scalable for high-resolution medical images. Experimental results demonstrate entropy values approaching the theoretical maximum (7.999), NPCR above 99.9%, and UACI up to 40%, confirming strong diffusion and resistance against differential and chosen-plaintext attacks.