PushCen-ADFL cuts communication 80% while boosting federated learning accuracy 6%

Asynchronous decentralized federated learning (ADFL) promises scalability by eliminating central coordination and global synchronization, but it suffers from excessive communication overhead, biased aggregation, and severe model drift—especially under non-IID data and directed topologies. To solve this, researchers Jiahui Bai, Hai Dong, and A.K. Qin introduce PushCen-ADFL, a novel framework accepted at KDD 2026. The approach couples communication, aggregation, and local stabilization in a shared centroid representation space. Clients exchange centroid-form messages, apply average-preserving push-sum mixing to correct aggregation bias, and use lightweight centroid regularization anchored in the same space to mitigate drift under heterogeneity and staleness. A bounded, sender-deduplicated buffer further improves robustness against irregular asynchronous arrivals.

Experiments on vision datasets demonstrate that PushCen-ADFL improves accuracy under data heterogeneity by up to 6% while reducing per-push communication cost by more than 80%, achieving a favorable accuracy-communication trade-off. The framework is designed for large-scale heterogeneous systems where peer-to-peer communication and client delays are common. By compressing messages into centroid representations and eliminating redundant sends, PushCen-ADFL makes asynchronous federated learning practical for real-world deployment, offering a path to stable, decentralized model training with minimal bandwidth overhead.