LPV-FL: A Lightweight Privacy-Preserving and Verifiable Federated Learning Scheme

In federated learning (FL), the aggregation server is typically assumed to be honest-but-curious–it faithfully follows the aggregation protocol but attempts to infer private information as much as possible. Adversaries may manipulate a dishonest server to maximize illegal benefits, such as reducing the number of aggregation rounds to save computation or fabricating aggregation results to affect model updates and performance. To ensure the security of FL aggregation, it is crucial to verify the correctness of the server-side aggregation results and maintain the consistency of model distribution during the FL process. To address the issues of high overhead and insufficient integrity in verifying dishonest servers, this paper proposes a lightweight privacy-preserving verifiable federated learning framework, named LPV-FL. This framework adopts a linear masking mechanism to protect gradient privacy, introduces a local verification method based on linear homomorphic hashing to improve verification efficiency, and incorporates blockchain to ensure the consistency and auditability of the global model. Experimental results show that LPV-FL reduces computation and communication overhead while guaranteeing integrity and verification performance. Compared with the baseline, it achieves a 63.18% reduction in verification time and a 98.73% reduction in communication overhead.