STDFL: A Spatio-Temporal-Aware Dynamic Federated Learning Framework for Spatial Crowdsourcing
Spatial Mobile Crowdsourcing (SMC) faces the dual challenge of ensuring privacy while managing dynamic, Non-IID spatio-temporal data. While Federated Learning (FL) offers a privacy-preserving solution, traditional aggregation suffers from severe model drift due to evolving spatio-temporal contexts. Furthermore, existing approaches often decouple prediction from scheduling, failing to translate predictive insights into tangible task allocation efficiency. To address these challenges, we propose STDFL, a prediction-driven dynamic framework tailored for SMC. It features a hierarchical architecture combining client-side lightweight models for micro-patterns and a server-side Transformer for global dependencies. To mitigate drift, our Dynamic Spatio-Temporal Perceiving Aggregation adaptively weights updates based on spatial similarity and temporal freshness. For privacy, we integrate client-level <inline-formula><tex-math notation="LaTeX">$(\epsilon, \delta )$</tex-math><alternatives><mml:math><mml:mrow><mml:mo>(</mml:mo><mml:mi>ε</mml:mi><mml:mo>,</mml:mo><mml:mi>δ</mml:mi><mml:mo>)</mml:mo></mml:mrow></mml:math><inline-graphic xlink:href="wang-ieq1-3673177.gif"/></alternatives></inline-formula>-Differential Privacy to ensure formal protection. Finally, we introduce a Prediction-Driven Scheduler (PDS) that leverages predictive potentials for bipartite matching, theoretical analysis and stress tests confirm PDS achieves linear scalability and zero policy training cost, offering a superior real-time deployment trade-off compared to RL approaches. Experiments on Chengdu and Nanjing datasets show STDFL significantly outperforms SOTA baselines in efficiency and fairness,while achieving near-centralized prediction accuracy under rigorous privacy guarantees and effectively bridging the utility–privacy trade-off in spatial crowdsourcing.