Personalized and differentially private federated learning for mobility data mining via the TrajectoryGuard framework
- 1Department of Information Systems and Cybersecurity, College of Computing and Information Technology, University of Bisha, Bisha, Saudi Arabia.
Abstract
This paper presents TrajectoryGuard, a personalized and privacy-preserving federated learning framework for secure and efficient mobility data mining. The framework integrates differential privacy, secure multiparty aggregation, and adaptive gradient compression to ensure that raw location traces, such as GPS logs or geo-tagged user activities, remain confined to user devices, thereby minimizing privacy leakage risks. Unlike conventional approaches, TrajectoryGuard introduces a dynamic noise-aware optimization mechanism and personalized aggregation scheme to enhance both model accuracy and robustness under non-IID and heterogeneous client scenarios. Experimental evaluation on the GeoLife dataset and the T-Drive Beijing taxi trajectory dataset demonstrates competitive performance, achieving 92.5% accuracy and 30% reduction in total per-round communication overhead (comprising gradient payload, model broadcast, and encryption metadata) while maintaining strong privacy guarantees (ε = 0.9 at σ2 = 1.0). Comparative results against FedAvg, FedProx, AdaptFL, and Fed-Ensemble confirm TrajectoryGuard’s superiority in convergence stability, privacy resilience, and predictive utility. The framework’s practical viability is illustrated through a real-world case study in urban trajectory forecasting, highlighting its relevance for applications in smart transportation and behavior modeling. The accuracy metric specifically corresponds to next-destination classification, a multi-class problem where the model predicts a destination category (e.g., home, workplace, or shopping location) from partial GPS trajectories. Evaluation follows an 80/20 time-based train/test split per user to prevent data leakage across temporal boundaries.
Related
- Hybrid machine learning framework for drought forecasting: integration of LSBoost, RQNN, CNN-RNN, CNN-LSTM and advanced data decomposition
- IceWatch: a multimodal deep-learning framework for monitoring glacial lake outburst floods, evaluated at Shisper Glacier, Karakoram
- Integrated machine learning and AHP-based framework for spatial delineation of rainwater harvesting potential
- A source-target transfer learning framework with physics priors for subsurface chlorophyll reconstruction in the Northwest Pacific Ocean
- SeisBench DAS: A machine learning framework for Distributed Acoustic Sensing
BibTeX
@article{alqazzaz2026personalized,
title = {{Personalized and differentially private federated learning for mobility data mining via the TrajectoryGuard framework}},
author = {Ali Alqazzaz},
journal = {PLOS One},
year = {2026},
month = oct,
doi = {10.1371/journal.pone.0358703},
url = {https://doi.org/10.1371/journal.pone.0358703},
note = {CC BY},
}
RIS
TY - JOUR
TI - Personalized and differentially private federated learning for mobility data mining via the TrajectoryGuard framework
AU - Ali Alqazzaz
PY - 2026
DA - 2026/10/29
JO - PLOS One
DO - 10.1371/journal.pone.0358703
UR - https://doi.org/10.1371/journal.pone.0358703
AB - This paper presents TrajectoryGuard, a personalized and privacy-preserving federated learning framework for secure and efficient mobility data mining. The framework integrates differential privacy, secure multiparty aggregation, and adaptive gradient compression to ensure that raw location traces, such as GPS logs or geo-tagged user activities, remain confined to user devices, thereby minimizing privacy leakage risks. Unlike conventional approaches, TrajectoryGuard introduces a dynamic noise-aware optimization mechanism and personalized aggregation scheme to enhance both model accuracy and robustness under non-IID and heterogeneous client scenarios. Experimental evaluation on the GeoLife dataset and the T-Drive Beijing taxi trajectory dataset demonstrates competitive performance, achieving 92.5% accuracy and 30% reduction in total per-round communication overhead (comprising gradient payload, model broadcast, and encryption metadata) while maintaining strong privacy guarantees (ε = 0.9 at σ2 = 1.0). Comparative results against FedAvg, FedProx, AdaptFL, and Fed-Ensemble confirm TrajectoryGuard’s superiority in convergence stability, privacy resilience, and predictive utility. The framework’s practical viability is illustrated through a real-world case study in urban trajectory forecasting, highlighting its relevance for applications in smart transportation and behavior modeling. The accuracy metric specifically corresponds to next-destination classification, a multi-class problem where the model predicts a destination category (e.g., home, workplace, or shopping location) from partial GPS trajectories. Evaluation follows an 80/20 time-based train/test split per user to prevent data leakage across temporal boundaries.
SN - 1932-6203
ER -