On the Push-Based Asynchronous Federated Learning: A Bias-Correction Aggregation Approach

In three linesPushCen-ADFL is an asynchronous decentralized federated learning framework reducing communication by 80% while improving accuracy by up to 6% under data heterogeneity. It employs shared centroid representation, average-preserving push-sum mixing, and lightweight centroid regularization to correct aggregation bias and mitigate model drift.

Summary generated by Claude — human-verified