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 through collaborative training, yet gradient leakage attacks can accurately reconstruct sensitive information from shared updates. While homomorphic encryption (HE) offers a formidable defense, its high computational and communication costs often make it impractical for real-world deployment. In this work, we propose FL SelHE Federated Selective Homomorphic Encryption, a framework that strategically secures only the most vulnerable layers of a model. By using a one-time sensitivity profiling based on gradient magnitudes, we identify critical layers and apply the CKKS encryption scheme selectively. Experimental evaluation demonstrates that the framework achieves 94.79 % accuracy, comparable to unencrypted training, while reducing the aggregation time by over 40 % compared to fully encrypted models. Furthermore, under gradient inversion attacks, the reconstructed images degrade to near-random noise (PSNR ≈ 5 dB and SSIM ≈ 0), confirming strong resistance to data leakage. These findings prove that selective encryption is a scalable and efficient solution for robust privacy in sensitive domains like healthcare.