Efficient verifiable secure aggregation protocols for federated learning

Abstract

Federated learning enables collaborative model training without direct access to clients’ local data, making it highly attractive for privacy-preserving analytics in resource-constrained environments. However, existing secure aggregation protocols remain vulnerable to privacy disclosure and malicious server tampering, and often incur substantial computational and communication overhead. In this paper, we propose a verifiable secure aggregation protocol that enables efficient aggregation in resource-constrained settings while guaranteeing the integrity of the aggregated results. Integrity of the aggregated result is guaranteed via the additive homomorphism of Shamir secret shares and a lightweight symmetric message-authentication code. Compared to VerifyNet, our protocol reduces aggregation overhead to only 1.25% of VerifyNet’s overhead, and under client dropouts it cuts RFLPV’s overhead by approximately 50%, while maintaining full privacy against semi-honest clients. Extensive simulations confirm that our method delivers strong security guarantees and operates efficiently under resource-constrained conditions, demonstrating its suitability for large-scale, dropout-prone federated learning deployments.