MedHE: Communication-Efficient Privacy-Preserving Federated Learning with Adaptive Gradient Sparsification for Healthcare

Federated learning (FL) enables privacy-aware training across distributed healthcare institutions without sharing raw patient data. However, standard FL remains vulnerable to membership inference attacks (MIA), and applying homomorphic encryption (HE) naively inflates communication by 5-10x, rendering it impractical over bandwidth-limited hospital links. We present MedHE, a framework that co-designs adaptive gradient sparsification with CKKS homomorphic encryption to simultaneously achieve cryptographic privacy and 97.5% communication reduction for healthcare NLP. MedHE transmits only the top-10% of gradient magnitudes packed into CKKS ciphertexts with error-feedback compensation, reducing communication from 1,277 MB to 32 MB. Across five independent trials, MedHE achieves macro-F1=0.950±0.005 and 89.5%±0.8% accuracy with no statistically significant difference from standard FL (p=0.32), while HE-only FL collapses to macro-F1=0.714 (±0.126), confirming sparsification is essential for stable predictions under encryption. Experimentally measured MIA resistance confirms multi-layer privacy: MedHE achieves 51.1% worst-case attack rate versus 56.3% for standard FL (ideal: 50%). Ablation confirms each component is essential: removing error-feedback drops F1 by 19.7%; removing batch packing inflates communication 13x; removing HE collapses privacy. MedHE provides a practical blueprint for privacy-preserving FL in bandwidth-constrained healthcare settings.