ADAPTIVE SENSITIVE-LAYERS BASED HOMOMORPHIC ENCRYPTION FOR SCALABLE AND SECURE FEDERATED LEARNING

Abstract

Federated Learning (FL) enables collaborative model training across decentralized clients while maintaining data locality; however, the protection of exchanged model updates remains a critical security challenge. While Full Homomorphic Encryption (FHE) ensures high confidentiality, its prohibitive computational and communication overhead often renders it impractical for resource-constrained FL systems. To address this bottleneck, we propose a Selective Homomorphic Encryption (SHE) framework that optimizes the utilityefficiency trade-off. Our approach leverages a one-time gradient sensitivity analysis to identify and encrypt only the most critical model layers using the CKKS scheme, while maintaining less sensitive layers in plaintext. Experimental evaluations conducted on the BloodMNIST dataset demonstrate that this selective strategy maintains a model utility nearly identical to standard FL. The proposed framework achieved an accuracy of 96.18% with a loss of 0.1651, performing comparably to the baseline FL (96.32%, loss : 0.1628) and outperforming the fully encrypted configuration (96.05%, loss: 0.1684). Furthermore, the selective scheme significantly improved system efficiency, reducing client computation time by 12.86% and cutting encryption-related overhead by up to 41.67%. These findings suggest that selective homomorphic encryption provides a highly practical and scalable solution for privacy-preserving FL deployments in sensitive domains such as healthcare and medical imaging.