Flower framework gets semi-async FL boost with FedSaSync strategy

A new research paper introduces FedSaSync, a semi-asynchronous federated learning strategy built on the Flower framework. Traditional synchronous FL waits for all clients to finish before aggregating, leading to idle time when some clients are slower (straggler effect). Fully asynchronous methods can drift in convergence. FedSaSync finds a middle ground: it allows partial synchronization, letting faster clients proceed without waiting for every straggler, while still maintaining enough coordination to ensure stable convergence. The implementation is a direct extension of Flower, making it accessible to existing FL deployments.

In experiments with heterogeneous clients (varying compute and data), FedSaSync showed improved robustness and significantly reduced idle time compared to fully synchronous baselines. The results demonstrate that semi-asynchronous FL can efficiently handle the straggler problem common in edge and distributed learning scenarios. This work, part of Víctor Hidalgo-Izquierdo's PhD thesis, provides a practical drop-in strategy for Flower users who need to scale FL across diverse devices without sacrificing convergence quality.