EVFLS: An Effective and Verifiable Federated Learning Aggregation Scheme

Federated learning collaboratively trains global models by sharing local gradients without exposing the original data. However, the sharing of gradients poses a significant threat to the privacy leakage of local data, highlighting the need for enhanced security measures in technological applications, attackers can compromise the integrity of the training dataset, whereas central servers have the capability to fabricate the aggregated results. The task of designing a verifiable and lightweight secure aggregation protocol remains a formidable challenge in the realm of technological advancements. In this paper, we propose an effective and verifiable federated learning security aggregation scheme to protect the privacy of data owners and verify aggregated results. To ensure the confidentiality of local data, we utilize a symmetric homomorphic encryption scheme enabling efficient masking of users’ local gradients while preserving their privacy. The central server conducts gradient aggregation on encrypted data and subsequently transmits the aggregated results to the task manager for decryption. Each user can efficiently verify the correctness of the aggregation results with homomorphic signatures. The final experiments yield qualitative comparisons and performance benchmarks, demonstrating that our scheme is capable of achieving a balance between secure aggregation and accuracy, integrity and efficiency. Thorough performance evaluations demonstrate the effectiveness of our proposed scheme in terms of minimal communication overhead and computational cost.