An Efficient and Secure Framework for Enhancing Data and Model Security in Healthcare AI

Abstract

The Artificial Intelligence (AI) enabled healthcare systems are require simultaneous defense of complex patient data and of the trained models. The paper proposes a dual-layer security architecture that combines AES-256 data encryption, homomorphic encryption (HE) for model protection, role-based access control (RBAC) with multi-factor authentication (MFA) and the digital watermarking for model ownership verification. The outline is evaluated on the MIMIC-III clinical dataset using a feedforward neural network for healthcare prediction. The experimental results show a test accuracy of 92.53%, data encryption and decryption times of 0.144s and 0.109s, model-level encryption and decryption times of 0.0019s and 0.0002s with a watermark tamper resistance of 99.95% under simulated adversarial attack. Compared with three recent baseline frameworks, the proposed approach achieves the highest tamper resistance and the fastest encryption or decryption operations the outperforms the layered cryptographic and access-control mechanisms as embedded into an AIML pipeline without the compromising low latency requirements of real world clinical deployment.