Privacy-Preserving Edge Intelligence Framework (PPEIF) Using Homomorphic Encryption and Knowledge Distillation for Efficient Electronic Health Records Management

The rapid integration of machine learning in healthcare emphasizes the need for privacy-preserving and efficient solutions, especially when managing sensitive Electronic Health Records (EHRs). Existing federated learning (FL) frameworks face significant challenges, including high communication overhead, computational inefficiency on resource-constrained edge devices, limited privacy guarantees during inference, and vulnerability to noisy or malicious updates. This study proposes a novel Privacy-Preserving Edge Intelligence Framework (PPEIF), designed specifically to overcome these limitations by combining Homomorphic Encryption (HE), Knowledge Distillation (KD), and an Attention-Based Aggregation Mechanism. In the PPEIF framework, a large teacher model trains lightweight student models via knowledge distillation, enabling efficient and encrypted inference directly at edge nodes. Homomorphic Encryption ensures that raw EHR data remains encrypted throughout local processing, preventing any data leakage. Instead of transmitting full model parameters, only encrypted distilled logits are shared, drastically reducing communication overhead. The attention mechanism dynamically weighs local contributions during global aggregation, mitigating the effects of noisy or malicious updates. Experimental evaluation on the MIMIC-III dataset demonstrates that PPEIF achieves 94.5% inference accuracy, significantly lower inference time (1.9 seconds), and up to 80% reduction in communication overhead compared to conventional FL methods. Privacy leakage risk is minimized, achieving the highest privacy level by protecting both model updates and inference results. Comparative analysis with state-of-the-art works further validates the superiority of PPEIF in terms of accuracy, scalability (supporting over 200 edge nodes), and real-time deployment feasibility. The proposed framework offers a robust and practical solution for secure, efficient, and scalable healthcare applications, setting a new benchmark for future privacy-preserving federated learning research in sensitive domains.