FL_SelHE: A Selective Homomorphic Encryption Based Federated Learning Framework for Robust Privacy Protection Against Gradient Leakage Attacks

Abstract

Federated learning (FL)promises to protect user data by training artificial intelligence (AI) models collaboratively without sharing raw information-yet gradient leakage attacks ruthlessly exploit shared model updates to reconstruct sen sitive personal details. While homomorphic encryption (HE) offers robust theoretical security, its crippling computational and communication costs limits its real-world adoption. To address this challenge, we propose FL_SelHE, a federated learning framework that applies selective homomorphic encryption guided by a one-time gradient sensitivity analysis. Parameters with high influence on model loss and strong encoding of identifiable patterns are encrypted using the CKKS scheme, while the remaining parameters remain unencrypted to preserve efficiency. This selective protection strategy achieves a balanced trade-off between privacy and performance, providing strong resistance against leakage attacks while significantly reducing encryption overhead compared to full HE without compromising security at critical points of exposure. The results underscore the potential of the proposed FL_SelHE to maintains strong predictive performance (94.79% accuracy) while lowering aggregation time by more than 40% compared to fully encrypted models. Against gradient inversion attacks, reconstructed images exhibit PSNR values around 5 dB and SSIM scores near zero, confirming that the selectively encrypted layers effectively suppress recoverable private content. These findings highlight the practicality of incorporating robust cryptographic privacy into federated learning through targeted protection rather than full encryption.