Statistical Heterogeneity

Statistical heterogeneity in federated learning (FL) arises from the diverse data distributions across participating devices, hindering the efficient training of a shared model. Current research focuses on mitigating this heterogeneity through improved aggregation techniques (e.g., FedAvg, FedProx), personalized model approaches, and adaptive client selection strategies that prioritize informative updates. These advancements aim to enhance model accuracy, robustness, and fairness in FL systems, impacting various applications, particularly those involving sensitive, decentralized data like medical imaging or industrial manufacturing. The ultimate goal is to enable effective collaborative learning while preserving data privacy and addressing the challenges posed by non-identically distributed data.