Context-Sensitive Differential Privacy Orchestration for Smart City Trajectory Data

Continuous data collection in smart city environments supports transportation systems, environmental monitoring, and real-time urban services through connected sensing infrastructures. However, these data streams may expose sensitive mobility patterns and contextual user information. This paper proposes a context-sensitive privacy orchestration (CSPO) framework that adapts local privacy budgets to heterogeneous spatial, temporal, behavioral, and activity density context. An entropy-based decision tree partitions the contextual space into disjoint leaves, and a fixed leaf-level allocation rule uses the contextual-risk score and mean density signal to assign bounded local privacy budgets. Protected coordinates are released with analytically calibrated Gaussian noise under a public bounded-location adjacency model. The evaluation uses a subset of the GeoLife GPS trajectory dataset, 500 fixed spatial-range queries, and five global privacy budgets. Across the evaluated operating points, CSPO produced lower query error than AdaTrace and PrivTrace while maintaining substantially lower runtime than AdaTrace. Analytic Gaussian Mechanism achieved slightly lower aggregate spatial-query error than CSPO, while CSPO provided context-sensitive heterogeneous local privacy budget allocation. The formal privacy guarantee applies to coordinate release under fixed contextual assignments and does not cover context construction or leaf assignment from raw sensitive data. A controlled component-wise ablation at εg = 1.0 showed that the complete CSPO configuration achieved the lowest Query MAE, MSE, and RMSE among the evaluated variants. These findings support context-sensitive privacy orchestration for the evaluated spatial-query workload within the stated adjacency model and experimental setting.