Unbiased Federated Learning for Heterogeneous Data Under Unreliable Links

Federated learning (FL) has emerged as a promising paradigm for privacy-preserving collaborative learning, enabling multiple devices to jointly train a global model without sharing their raw data. However, the bias in FL training significantly reduces its performance. This poster presents a novel FL algorithm to counteract bias for performance improvement. First, we provide a global perspective for analyzing the causes of bias, data heterogeneity and transmission probability. Then, we propose a method that introduces a regularized local training method and a reweighted aggregation strategy to jointly mitigate bias. Through extensive experiments on real-world datasets, we demonstrate that our method outperforms various baseline FL methods in terms of convergence speed and accuracy.