Med Fusion Crypt: Cross-Modal Attention-Guided Selective Encryption and Pairing-Free Proxy Re Encryption for Secure Multimodal Medical Image Sharing in IoMT

Abstract

The rapid growth of the Internet of Medical Things (IoMT) and telemedicine has increased the exchange of multi-modal medical images such as CT, MRI, and PET across hospitals, cloud servers, and specialists. These images carry highly sensitive patient information and are vulnerable to interception, tampering, and unauthorized redistribution. Existing blockchain-based proxy re-encryption (PRE) schemes mostly target textual electronic medical records, rely on computationally expensive bilinear pairings, and treat each imaging modality independently, which leads to redundant encryption, high latency, and poor suitability for resource-constrained IoMT devices. To address these limitations, this paper proposes a privacy-aware multi-modal medical image transmission framework that integrates a novel deep learning mechanism with a blockchain-based lightweight re-encryption scheme. First, a Cross-Modal Attention Fusion and Region-of-Interest Network (CMAF-RoINet) fuses complementary features from multiple imaging modalities into a compact representation and automatically localizes diagnostically critical regions. This enables selective encryption, so that high-strength protection is applied to sensitive regions while computational cost is reduced for background areas. Second, a deep learning driven chaotic key-stream generator, combining an LSTM network with a hyper-chaotic map, produces image specific, non-repeating keys that resist differential and brute-force attacks. Third, a pairing-free, elliptic-curve based unidirectional proxy re-encryption scheme re-encrypts only the symmetric session key rather than the full image. This allows patients to delegate access to authorized clinicians without exposing plaintext to the proxy and resists collusion between the proxy and revoked users. Finally, a consortium blockchain with smart contracts manages patient consent, re-encryption key issuance, and revocation, and records an immutable audit trail, while encrypted images are stored off-chain in the InterPlanetary File System (IPFS) with only their content hashes stored on-chain for integrity verification. The proposed framework is evaluated on publicly available multi-modal datasetlikeBraTS. Experimental results show that the scheme achieves an information entropy, NPCR, UACI, and near-zero adjacent pixel correlation, with lossless reconstruction (PSNR of dB, SSIM) for authorized receivers. Compared with existing pairing-based PRE and blockchain image encryption schemes, the proposed approach reduces re-encryption computation cost and end-to-end transmission latency, while achieving a blockchain throughput. Security analysis further demonstrates resistance to chosen-plaintext, collusion, replay, and impersonation attacks. These results indicate that the proposed framework offers a secure, efficient, and patient-centric solution for multi-modal medical image sharing in IoMT-enabled healthcare systems.