Privacy Protection Bottom-up Hierarchical Federated Learning with Class Imbalanced Data

Jing Zhang, Siqi Li, Jiajia Li, Xiufeng Xia, Yiping Teng, Anzhen Zhang

Abstract: Federated learning (FL) is a distributed machine learning method that enables multiple participants to contribute a well performed global model while their private training data remains in local devices. FL is promising in the edge computing system which has a large corpus of decentralized data and requires data privacy. However, traditional FL algorithms perform poorly with not independently and identically distribution data, especially highly skewed class imbalanced datasets. When solving class imbalance problems in FL, it is necessary to have prior knowledge of data distribution information, which cannot protect data distribution privacy. To fully protect privacy, we build a privacy protection bottom-up hierarchical federated learning (FedPBH) framework, which alleviates the imbalances by 1) Data sampling based on global data distribution, and 2) Bottom-up client participation. The proposed framework relieves global imbalance by data sampling based on the global data distribution which is obtained through privacy protection collaborative data distribution evaluation. For averaging the local imbalance, the proposed method creates bottom-up client participation, and these clients in the same local server asynchronously train their models. Experiments demonstrate that our FedPBH model provides full privacy protection with high classification performance.