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.

Keywords

MSC

68T05, 68P25, 94A60

Cite this article

@article{cai-8521,
  author  = {Mohammed El Aymene Beyat and Mohammed Kamel Benkaddour and Abdessalam Moualdi and Abdelalim Belmesmar and Mohammed El Amine Beyat},
  title   = {FL{\_}SelHE: A Selective Homomorphic Encryption-Based Federated Learning Framework for Robust Privacy Protection Against Gradient Leakage Attacks},
  journal = {Computing and Informatics},
  volume  = {45},
  number  = {5},
  pages   = {1236--1271},
  year    = {2026},
  issn    = {1335-9150},
  doi     = {10.31577/cai_2026_5_1236}
}
Download .bib