OVP-FL: Outsourced Verifiable Privacy-Preserving Federated Learning

Abstract

Federated learning, a prominent method, eliminates the need for users to upload their original data, enabling collaborative model training through the transmission of only the gradient information of their models. However, a deeper exploration of federated learning has uncovered vulnerabilities wherein the gradient information uploaded by users can be exploited by adversaries to reconstruct users' original data. Additionally, ensuring the integrity of the aggregation result remains a primary focus of research to protect users' legitimate interests. To address these issues simultaneously, this study proposes a new framework called Outsourced Verifiable Privacy-Preserving Federated Learning. This framework aims to provide reliable privacy protection to users. Additionally, it includes a validation function to detect malicious aggregation models submitted by the server, providing an almost cost-free solution that accommodates the possibility of dropout. Finally, the paper concludes with a comprehensive security analysis that evaluates the reliability of the scheme using different datasets. In comparison to VerifyNet, our scheme demonstrates a significant advantage, with an approximate 100x improvement in the overall performance. And, the additional drop overhead is negligible. Simulation experiments demonstrate a significant improvement, including a reduction in communication and computation costs, showcasing the efficacy compared to existing verifiable privacy-preserving federated learning methods.