Fed-SecureCvT: optimized federated learning with cryptographic protection and transformer-based intrusion detection for HER

Digital healthcare technologies have transformed the sharing of medical information and collaborative clinical decision-making while raising critical concerns about privacy leakage, unauthorized data manipulation, and cyber risks in remote healthcare settings. Current methods generally focus on collaborative learning, cryptographic security, and network security as separate functions, which reduces their potential to offer an all-round solution in an integrated framework and leaves a key gap in research for the context of secure multi-institutional healthcare systems. To address these challenges, this work presents a multilayer secure collaborative framework that combines intelligent learning, optimized cryptography, integrity assurance, and intrusion detection in a seamless coordinated platform for secured EHR management. The proposed approach is Federated Learning and Federated Dynamics optimization (FL–FedDyn), which could support decentralized medical prediction across multiple hospitals without sharing raw patient data, thus protecting privacy and enhancing the generalization performance of the medical prediction model under heterogeneous data distributions. A Decentralized Multi-Authority Attribute-Based Encryption (DMA-ABE) mechanism is embedded to promote fine-grained access control, with each authority independently providing credentials and eliminating the single-point trust dependency. The Mandrill Optimization Algorithm (MOA) is used to generate optimized keys, which maximizes the entropy and minimizes the computational cost to enhance the cryptographic robustness and efficiency. The Merkle Directed Acyclic Graph (Merkle-DAG) structure is used to link encrypted records via hash relationships for efficient tamper detection and secure storage. Moreover, an Intrusion Detection System (IDS) based on a Convolutional Vision Transformer (CvT) is incorporated to constantly monitor the traffic in a healthcare network and identify known and unknown threats through hybrid feature learning of spatial and attention features. The modules are executed in a streamlined process that includes data preprocessing, collaborative training, encryption, secure storage, integrity validation, and real-time threat monitoring, all of which work together to provide holistic protection over multiple layers of security. Experimental evaluation conducted on the EHR and CIC-IDS2017 datasets demonstrates that the proposed framework achieves superior performance with a prediction accuracy of 99.31%, convergence time of 118.42 s using only 36 communication rounds, privacy preservation level of 98.74%, encryption time of 5.26 ms, integrity verification time of 11.38 ms, and intrusion detection accuracy of 99.31% with a false alarm rate of 1.38%. The framework also demonstrates efficient resource utilization with a reduced CPU consumption of 54.18% and energy usage of 0.196 mJ compared with existing techniques. These results confirm that the proposed multilayer approach provides an effective, scalable, and secure solution for collaborative healthcare data management in distributed medical infrastructures.