Privacy-Preserving Yet Vulnerable: Data Poisoning Attacks Against Differential Privacy Sparse Mobile Crowdsensing System
IEEE Transactions on Mobile Computing · 2026-07 · 25 (7) · 9594-9611
DOI: 10.1109/tmc.2026.3655365 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Mobile Computing, 2026-07. Verfasst von Chengxin Li und 5 weiteren Autor:innen. Die Arbeit wurde bisher 2-mal zitiert (Crossref) und nennt 59 Literaturangaben. Gefördert von: National Natural Science Foundation of China International Cooperation Key Project, National Natural Science Foundation of China.
Abstract (Originalsprache)
Sparse Mobile CrowdSensing (SMCS) reduces sensing costs while preserving data quality, shedding a new paradigm on data collection. However, the reliance on user-contributed data raises significant privacy concerns. While Local Differential Privacy (LDP) offers strong theoretical guarantees for protecting user privacy, the noise it introduces not only degrades inference accuracy but also creates opportunities for new security threats, in particular data poisoning attacks. In this work, we study the vulnerability of differentially private SMCS systems under data poisoning. We propose a novel attack framework formulated as a bi-level optimization problem with spatiotemporal constraints. Specifically, the lower-level problem performs robust data inference using differentially private matrix factorization with a mixture of Gaussians, enabling accurate reconstruction from noisy and incomplete data. The upper-level problem, from the attacker’s perspective, aims to maximize inference error while maintaining stealth, which is achieved by enforcing temporal and spatial similarity between poisoned and normal data. To solve the bi-level problem, we develop an alternating algorithm that combines projected gradient ascent with the Expectation-Maximization procedure to obtain the optimal poisoning strategy. Extensive experiments on real-world datasets show that attackers can effectively degrade the system’s inference performance while concealing their attack behavior. Additionally, we observed a positive correlation between attack intensity and spatiotemporal constraint loss, enabling attackers to adapt their strategies to specific deployment requirements. These findings highlight the structural vulnerability of differential privacy SMCS systems to data poisoning attacks.
Förderung
- National Natural Science Foundation of China International Cooperation Key Project
- National Natural Science Foundation of China
Zitieren
APA
Li, C., Gu, Y., Qiao, P., Pan, S., Sakurai, K., & Li, Z. (2026). Privacy-Preserving Yet Vulnerable: Data Poisoning Attacks Against Differential Privacy Sparse Mobile Crowdsensing System. IEEE Transactions on Mobile Computing, 25(7), 9594-9611. https://doi.org/10.1109/tmc.2026.3655365
Vancouver
Li C, Gu Y, Qiao P, Pan S, Sakurai K, Li Z. Privacy-Preserving Yet Vulnerable: Data Poisoning Attacks Against Differential Privacy Sparse Mobile Crowdsensing System. IEEE Transactions on Mobile Computing. 2026;25(7):9594-9611. doi:10.1109/tmc.2026.3655365
BibTeX
@article{Li2026,
author = {Li, Chengxin and Gu, Yujie and Qiao, Pengpeng and Pan, Shengli and Sakurai, Kouichi and Li, Zhetao},
title = {Privacy-Preserving Yet Vulnerable: Data Poisoning Attacks Against Differential Privacy Sparse Mobile Crowdsensing System},
journal = {IEEE Transactions on Mobile Computing},
year = {2026},
volume = {25},
number = {7},
pages = {9594--9611},
doi = {10.1109/tmc.2026.3655365}
}
EUVIMED