Adaptive Sensitive-Layers Based Homomorphic Encryption for Scalable and Secure Federated Learning

Resumen

Federated learning (FL) allows several clients to train a shared model without moving their data to a central server, but protecting the exchanged model updates is still a major concern. Fully homomorphic encryption (HE) provides strong confidentiality, although its computation and communication costs can be too high for practical FL systems. To reduce this cost, we propose a selective homomorphic encryption framework that applies CKKS-based encryption only to the layers marked as most sensitive through a one-time gradient sensitivity analysis. The remaining layers stay unencrypted, which cuts down processing time while still offering meaningful privacy protection. Experiments on the BloodMNIST dataset show that this selective approach keeps the model very close to standard FL. In our tests it reached 96.18% accuracy with a loss of 0.1651. The baseline FL run reached 96.32% accuracy with a loss of 0.1628, while the fully encrypted setup reached 96.05% accuracy with a loss of 0.1684. The selective scheme also reduced client computation time by 12.86% and cut encryption-related overhead by up to 41.67%. These results suggest that selective homomorphic encryption gives a useful trade-off between privacy and efficiency, so it can be used in real FL deployments in sensitive areas such as healthcare.