A Lightweight Privacy-Preserving and Verifiable Federated Learning-Based Protocol

Federated learning is a combination of distributed technology and machine learning technology, which does not require model training participants to share their original data, but only requires these participants to share their local models (or gradients) with a third party for aggregation. However, existing researches suggest that shared original models (or gradients) may also lead to privacy breaches for some participants. Some third parties performing aggregation tasks may also return falsified aggregation results to the participants. Thus, federated learning still faces many security challenges. In this paper, we propose a lightweight privacy-preserving verifiable federated learning-based protocol. Specifically, we protect the original model parameters by splitting the model and masking the split model with a mask, which effectively protects the original model parameters even in the case of collusion among multiple entities. In addition, we design a lightweight dropout handling scheme and a verification scheme to allow the participants to quickly verify the aggregation results from the server, which reduces the harmful effects of erroneous model updates on the entire federated learning system. Finally, we test our proposed scheme on several popular public datasets and compare it with several schemes. The experimental results show that our scheme can guarantee high accuracy and verification efficiency even while providing privacy-preserving and verifiable aggregation results.