Work

SHELF+: A hierarchical homomorphic encryption-based framework for secure and efficient federated learning

Year: 2026
Type: article
Abstract: In this paper, we introduce SHELF+, an efficient and scalable federated learning framework that integrates homomorphic encryption (HE) with hierarchical aggregation and various system-level optimizations. HE-based federated learning methods often face significant computational and memory challenges. In contrast, SHELF+ introduces a clustered aggregation scheme, integrated with Top- K Sparsification, Grouped Packing, and Quantization techniques, to significantly reduce these overheads while maintaining encrypted aggregation. We conduct extensive experiments across MNIST and two medical imaging benchmarks from MedMNIST (PneumoniaMNIST and OrganAMNIST). The results show that SHELF+ achieves up to a 9.2 × reduction in memory usage and around 79.8% decrease in computation time relative to SHELF, while maintaining competitive model accuracy. An ablation study further examines the contribution of each optimization technique. The findings indicate that SHELF+ lowers the overhead of encrypted federated aggregation while allowing the efficiency–accuracy trade-off to be adjusted for different client scales and task complexity.
Language: English
Open Access status: gold
Estimated APC paid: $1,990
OpenAlex ID: W7236865320
FWCI:
Cites: 25
Cited by:
Related to: 10