SAEV: Secure Aggregation and Efficient Verification for Privacy-Preserving Federated Learning

Federated learning (FL) emerges as a promising paradigm, relentlessly pursuing excellence in efficiency, privacy preservation, and security-the holy trinity that underpins the fundamental philosophy and practical implementation of FL. However, previous scholarly works have predominantly focused on the privacy protection and secure aggregation in the realm of FL. Undoubtedly, efficiency still plays a pivotal role in determining FL's viability in real-world applications. To improve efficiency while ensuring privacy protection and secure aggregation, this work proposes a verifiable privacy-preserving federated learning framework tailored for practical applications. Firstly, a novel aggregation rule, constrained M maximum security aggregation, forces the server to securely aggregate the local gradients from M users without relying on any auxiliary servers, thereby considerably decreasing the communication overhead. Secondly, regardless of the dimension of the aggregated gradient being verified, our scheme performs a single verification per user per round. Through security analysis, our scheme could guarantee some given security requirements. Besides, extensive experiments show that under the proportion that the gradient dimension $(d)$ to the number of users $(n)$ is 1:2 and 2:1, ours is $\times 1.3$ and $\times 1$ faster for total runtime and is $\times 3$ and $\times 7$ lower for total communication cost compared with a state-of-the-art framework, respectively. Therefore, our FL framework is more suitable for practical application in real life.