Staleness-aware adaptive asynchronous federated learning for secure spatial crowdsourcing

Spatial crowdsourcing (SC) platforms increasingly depend on-device predictive models for critical services such as estimated time of arrival (ETA) prediction, demand forecasting, and task allocation. While asynchronous federated learning (AFL) mitigates the straggler effect, enabling collaborative model training on heterogeneous, privacy-sensitive data without centralising raw information, its real-world application in SC is hindered by gradient staleness and unbalanced client participation. These effects often degrade global model convergence and compromise both fairness and robustness. This paper proposes CrowdFedAsync, a staleness-aware adaptive AFL framework for secure SC systems. The proposed framework introduces dual staleness weighting that jointly accounts for delay-based and divergence-based staleness, along with Representative Update Buffering (RUB), which triggers aggregation only when buffered updates satisfy representativeness constraints across heterogeneous regions and device groups. By reducing the influence of outdated updates and preventing dominance by high-frequency contributors, CrowdFedAsync stabilises learning under volatile participation conditions. Extensive experiments on representative mobility and location-centric datasets, including NYC Taxi, Yelp, Geolife, and Foursquare, show that CrowdFedAsync improves average predictive utility by about 8.5% over FedAvg, reaches target performance faster than synchronous, semi-asynchronous, and asynchronous baselines, achieves better worst-group robustness with smaller group disparity, and remains more resilient under increasing staleness than delay-only and fixed-size buffered alternatives. These findings suggest that jointly modelling update usefulness and representativeness is an effective design principle for federated models in IoT-driven SC systems.