MobiTask: A Federated Learning-Based Task Migration Strategy for Mobile Crowdsensing
Siyuan Yin, Haifeng Jiang, Chaogang Tang, Shuo Xiao, Huaming Wu · IEEE Transactions on Computational Social Systems · 2025
In mobile crowdsensing (MCS), the quality of sensing is often enhanced through the rational allocation of tasks in the initial phase. However, during task execution, some participants may fail to complete their tasks on time, adversely affecting the overall completion rate. Although existing studies have introduced the concept of task migration, they lack a comprehensive migration strategy. To address this issue, we propose MobiTask, a federated learning (FL)-based task migration strategy for MCS. The proposed strategy consists of two stages: task execution progress prediction and task successor selection. In the task execution progress prediction stage, various factors leading to task delay are considered. A long short-term memory (LSTM) model with an attention mechanism is designed to predict participants’ task execution progress. Based on the predictions, participants who are unlikely to complete tasks on time are identified and removed. In the task successor selection stage, a multiagent reinforcement learning (MARL) algorithm incorporating a graph attention network (GAT) is proposed to construct a candidate pool and select appropriate task successors. Additionally, FL is employed to train the models, mitigating the impact of device heterogeneity while ensuring strategy security and stability. Simulation results demonstrate that MobiTask outperforms existing baseline strategies in terms of task completion rate and migration cost, validating its effectiveness and feasibility.