The proliferation of machine learning services in mobile crowd sensing (MCS) brings both new opportunities and severe privacy challenges. Location leakage, identity linkage, and centralized trust bottlenecks remain critical risks, while conventional differential privacy (DP) techniques often degrade service responsiveness due to noise generation latency and increased training overhead. To address these issues, this paper proposes PMLSF, a Privacy-Preserving Machine Learning Service Framework that integrates differential privacy, deep reinforcement learning (DRL), and blockchain-based orchestration to enable secure, efficient, and scalable task allocation in MCS environments. Specifically, we design Optimized Geo-indistinguishability (OptGI), a lightweight DP mechanism that reduces noise generation latency while maintaining rigorous privacy guarantees. We further introduce TopK Double Deep Q-Network (TopK-DDQN), a task allocation service that prunes infeasible actions and supports latency-aware decision-making over privatized inputs. To eliminate single points of failure, PMLSF employs a permissioned blockchain orchestration layer, ensuring anonymity, accountability, and distributed trust without imposing prohibitive overhead. Extensive experiments demonstrate that PMLSF achieves low latency privacy protection and significantly improves allocation utility and profit compared with state-of-the-art methods, positioning it as a scalable foundation for integrating DP methodologies into machine learning-driven MCS services.
Li et al. (Thu,) studied this question.