Decentralized Reputation-based Leader Election for Privacy-preserving Federated Learning on Internet of Things

The Internet of Things (IoT) based on Deep Learning (DL) technologies has facilitated people's lives in various aspects. However, since the IoT data used to train DL models contains users' sensitive personal information, privacy and security concerns arise in IoT. Although many works have presented security solutions for the privacy and security concerns on IoT, they cannot monitor the model quality of data owners, resulting in unusable models making misleading decisions, and they cannot defend against the curious participants inferring private data in the model training process. In this paper, we propose a Decentralized Reputation-based Leader Election scheme (DeRLE) for privacy-preserving distributed model training in IoT based on Federated Learning (FL) and Blockchain. DeRLE adopts decentralized model training while preserving privacy. To avoid a single point of failure in FL, DeRLE elects a refreshed leader in each epoch of model update, which prevents the fixed server of the basic FL from deriving sensitive data from historical local models. Furthermore, to protect the privacy of local models in FL, we design a reputation-based Differential Privacy (DP) mechanism to supervise the quality of local models and encourage data owners to inject reasonable DP noise. We conduct extensive experiments using Hyperledger Fabric and MNIST. The evaluation results confirm the fairness of DeRLE's probability distribution and demonstrate its feasibility and effectiveness.