FedPSA: Modeling behavioral staleness in asynchronous federated learning

Asynchronous Federated Learning (AFL) has become an important research direction in recent years. By avoiding waiting for slow clients and enabling concurrent training, AFL accelerates training compared with standard federated learning. However, asynchrony inevitably introduces stale updates, which may degrade model performance. Existing methods mostly adopt round difference to quantify staleness, a coarse-grained indicator that ignores intrinsic model characteristics and restricts the overall performance. In this work, we propose FedPSA, a fine-grained asynchronous federated learning framework. Specifically, we measure model obsolescence via parameter sensitivity approximated with second-order Taylor expansion and the empirical Fisher diagonal on a shared calibration batch. We further maintain a dynamic momentum queue to track recent update magnitudes, infer the real-time training stage, and adaptively adjust temperature-controlled aggregation weights. Extensive experiments on MNIST, FMNIST, CIFAR-10, and CIFAR-100 under IID and non-IID settings demonstrate that FedPSA generally outperforms existing methods in terms of final accuracy and convergence, achieving up to 6.37% improvement over baselines and 1.93% over state-of-the-art approaches.