Abstract
Mobile crowd sensing (MCS) frequently relies on truth discovery to aggregate noisy, conflicting reports into reliable estimates. Existing approaches often either risk exposing user data or overlook heterogeneous privacy needs and task-specific reliability, limiting aggregation fidelity. This study presents a task-wise, personalized, privacy-preserving truth discovery framework that learns per-user, per-task weights to enable high-quality aggregation while protecting both location and data privacy. Structural privacy is realized via aggregate-only Paillier homomorphic encryption-only aggregate sums are decrypted at the cloud-and task-scoped unlinkable pseudonyms that prevent cross-task linkage. The design also supports fine-grained incentives, aligning rewards with task-level contributions without revealing raw readings or identities. Evaluations on real-world MCS temperature traces and simulated workloads show accuracy relative to a non-private baseline (MAE/RMSE on the order of 10-5), fast and stable convergence under a uniform stopping rule, and predictable scaling with users, tasks, and key sizes; cloud-side decryption is the dominant cost, whereas the iterative solver remains stable. Overall, personalized weighting combined with structural privacy delivers practical, high-quality aggregation for privacy-critical MCS deployments.