An Efficient Privacy-Preserving Secure Aggregation Scheme for Federated Learning with Input Verification and Dropout Resistance

Federated learning, as a distributed machine learning paradigm, allows multiple participants to collaboratively train a shared model without sharing their local data. However, the increasing demand for privacy protection during data aggregation within distributed systems underscores the persistent challenge of ensuring both security and efficiency. Many existing Privacy-Preserving Machine Learning (PPML) schemes relying on homomorphic encryption introduce substantial computational overhead during aggregation, rendering them impractical for large-scale PPML applications involving resource-constrained participant devices. Moreover, device dropout events and data poisoning attacks perpetrated by malicious clients adversely affect the integrity of the aggregated results. To address these challenges, this paper proposes an efficient privacy-preserving secure aggregation scheme capable of tolerating participant dropout at arbitrary stages and securing data against both semi-honest and malicious participants. By integrating input verification protocols and applying gradient masking techniques, the scheme enhances its resilience against malicious attacks while ensuring user data privacy. Leveraging the additive homomorphic property of Shamir's secret sharing enables efficient global mask recovery, significantly optimizing the scheme's efficiency. Experimental results demonstrate that the proposed scheme significantly outperforms baseline methods in computational efficiency, communication overhead, and security robustness. By effectively balancing high privacy protection with practical feasibility, this scheme presents a promising solution for secure multi-party aggregation in large-scale distributed systems.