Tackling System Induced Bias in Federated Learning: Stratification and Convergence Analysis

In federated learning, clients cooperatively train a global model by training local models over their datasets under the coordination of a central server. However, clients may sometimes be unavailable for training due to their network connections and energy levels. Considering the highly non-independent and identically distributed (non-IID) degree of the clients’ datasets, the local models of the available clients being sampled for training may not represent those of all other clients. This is referred as system induced bias. In this work, we quantify the system induced bias due to time-varying client availability. The theoretical result shows that this bias occurs independently of the number of available clients and the number of clients being sampled in each training round. To address system induced bias, we propose a FedSS algorithm by incorporating stratified sampling and prove that the proposed algorithm is unbiased. We quantify the impact of system parameters on the algorithm performance and derive the performance guarantee of our proposed FedSS algorithm. Theoretical and experimental results on CIFAR-10 and MNIST datasets show that our proposed FedSS algorithm outperforms several benchmark algorithms by up to 5.1 times in terms of the algorithm convergence rate.