LO-GDRL: Privacy-preserving online task allocation based on Lyapunov optimization and graph-based deep reinforcement learning in mobile crowdsensing
Yuhong Tan, Tao Peng, Guojun Wang, Qin Liu, Tian Wang · Computer Networks · 2026
In Mobile Crowdsensing (MCS), online task allocation ensures timely task completion and improves overall system performance in dynamic environments through real-time scheduling and optimizing resource utilization. Existing Deep Reinforcement Learning methods have several limitations, including poor model performance, low system stability, and the problem of data privacy leakage. To address these issues, this paper proposes a lightweight privacy-preserving online task allocation framework called LO-GDRL (Lyapunov Optimization with Graph-based Deep Reinforcement Learning). LO-GDRL formulates NP-hard online task allocation as a graph-constrained optimization problem and designs a Deep Reinforcement Learning method with a new Dual-branch Graph Attention Dueling Network to enhance dynamic environment adaptation and complex dependency capture capability. To improve system stability, LO-GDRL establishes a dynamic-queue mechanism for dynamic resource coordination based on Lyapunov Optimization. Additionally, while preserving worker location privacy through differential privacy, the system achieves an optimal privacy-performance trade-off. The case studies on the simulation data of MCS systems based on the two real-world datasets verify the effectiveness of the proposed framework and demonstrate that our framework achieves more stable and superior performance across diverse environments compared to state-of-the-art methods.