Abstract
Crowdsensing systems face significant challenges in maintaining long-term user participation while protecting individual privacy. This paper introduces LEPA (Long-tErm Privacy-preserving data Aggregation), an online incentive mechanism designed to balance data accuracy, privacy costs, and user retention. By integrating Local Differential Privacy (LDP) with Lyapunov optimization, LEPA transforms long-term participation constraints into a stable queue problem, ensuring users remain active through fair selection and compensation. Experimental results demonstrate that LEPA maintains nearly 100% user retention compared to traditional static models.
Introduction
Mobile crowdsensing (MCS) has become a cornerstone for applications like traffic monitoring (e.g., Waze) and environmental sensing. However, two major hurdles prevent these systems from scaling effectively: privacy concerns and user retention. Users are often hesitant to share raw GPS or sensor data that could reveal sensitive habits. Furthermore, most existing incentive models are "static," focusing on single-round snapshots. If a user is not selected frequently enough, they fail to earn sufficient cumulative compensation to justify the battery and data drain, leading to app uninstallation or "churn."
System Architecture
The LEPA framework consists of a Fusion Center (FC) and a set of Participatory Users (PUs). The FC manages a task pool and employs an incentive mechanism to select winners based on their bids, which reflect both sensing and privacy costs.
As shown in the architecture, the process involves a multi-step cycle: the FC requests tasks, PUs submit bids, the incentive mechanism selects a winner set and issues payments, and finally, PUs provide perturbed data for aggregation.
Methodology
To address the "strategy coupling" problem—where a selection decision today affects participation tomorrow—the authors propose several key techniques:
- Local Differential Privacy (LDP): Instead of trusting a central server, users add "well-calibrated" Laplacian noise to their data locally. The paper uses tail bound theory (Chebyshev’s inequality) to quantify how this noise impacts aggregation accuracy, bypassing complex non-integrable distributions.
- Lyapunov Optimization: To manage long-term participation, the authors maintain a "virtual request queue" for each user. This queue grows when a user is not selected, naturally forcing the system to prioritize neglected users to ensure they earn enough to stay active.
- Reverse Combinatorial Auction: The mechanism is designed to be Truthful (bidding true cost is the optimal strategy) and Individually Rational (users never lose money). The online auction has a computational complexity of $O(N^3 + N^2M)$, making it suitable for real-time applications.
Evaluation and Results
The paper evaluates LEPA through extensive simulations involving 100 to 200 users and 10 concurrent tasks. The primary metric for success is the ability to maintain a stable user base over time.
As illustrated in the results, a standard static auction sees more than 50% of users drop out within 100 time slots. In contrast, under the LEPA mechanism (referred to as the Online LPRC auction in simulations), the user base remains almost 100% intact. Additionally, LEPA was found to be more cost-efficient than compulsory selection models, significantly reducing total payments by the Fusion Center while maintaining high data accuracy.
Conclusion
LEPA provides a robust solution for sustainable crowdsensing by aligning system-wide accuracy goals with individual privacy and economic needs. By treating user retention as a stability problem, it ensures the long-term viability of participatory sensing networks. Future work may explore heterogeneous privacy requirements where different users set unique privacy "price tags."
Summary updated