The Accuracy-Privacy Trade-off of Mobile Crowdsensing

Mohammad Abu Alsheikh, Yutao Jiao, Dusit Niyato, Ping Wang, Derek Leong, Zhu Han

Abstract

Mobile Crowdsensing (MCS) utilizes everyday devices as a distributed sensor network, yet it is often hindered by the "Privacy Paradox": users desire privacy, while providers require high-quality data. This paper introduces a novel incentive framework that treats privacy as a core economic variable. By implementing a Vickrey-Clarke-Groves (VCG) auction mechanism and $k$-anonymity coalitions, the system pays users based on their marginal contribution to model accuracy. Experimental results on a real-world activity recognition dataset demonstrate that this approach effectively balances the trade-off between user privacy and system performance, identifying "critical points" where excessive anonymization renders data counterproductive.

Introduction

Mobile crowdsensing (MCS) is a powerful paradigm where everyday devices—smartphones, wearables, and IoT gadgets—act as a distributed sensor network. While this "crowd intelligence" can power everything from real-time traffic maps to public health monitoring, it faces a fundamental roadblock: The Privacy Paradox. Users want to keep their personal habits private, but service providers need high-quality, "true" data to ensure their analytics are accurate.

Existing incentive models for MCS often focus purely on monetary compensation, ignoring the inherent "cost" of privacy. If users feel their privacy is at risk, they may provide fake data or opt out entirely. Conversely, if users protect their privacy by adding excessive noise (anonymization), the resulting analytics may become useless. This research proposes a truthful incentive mechanism that pays users based on their "marginal contribution" to a model's accuracy rather than just the volume of data provided.

The Privacy-Aware Incentive Framework

The proposed framework introduces a neutral mediator to manage the interaction between users and service providers. The system relies on three core components:

  1. Marginal Contribution Payoff: Using Vickrey-Clarke-Groves (VCG) auction logic, a user’s payoff is calculated as the difference between the model's accuracy with their data and the accuracy without it. "Pivotal users" who significantly boost accuracy receive higher rewards, while those who submit noisy data that hurts the model may be penalized.
  2. Privacy Levels ($t$-closeness): Users choose their own privacy level by adding Gaussian noise to their data. While higher noise variance increases privacy, it typically reduces the potential payoff by lowering the data's contribution to model accuracy.
  3. Crowdsensing Coalitions: To further protect identity, users can form coalitions to achieve $k$-anonymity. They pool data and submit it under a single identity. The group's total payoff is then distributed among members using the Shapley value, ensuring fair compensation based on individual contributions to the group's data quality.
Architecture of the mediator-based auction framework, illustrating user coalitions and data anonymization processes.

Experimental Evaluation

The framework was evaluated using a real-world activity recognition dataset involving 36 users performing six activities (walking, jogging, sitting, standing, upstairs, and downstairs). Accelerometer data was collected at 20Hz, totaling over 1 million samples. A deep learning model was employed to classify these activities, serving as the primary service provided by the platform.

Key Results and Insights

The study revealed several critical insights into the dynamics of crowdsensing:

  • Quality Over Quantity: There is no direct correlation between the volume of data provided and the resulting accuracy boost. A small amount of high-quality, "clean" data can be more pivotal—and thus more valuable—than massive amounts of noisy data.
  • The "Critical Point": There is a threshold of anonymization beyond which a user’s data becomes a liability. The researchers identified "critical points" where the service provider is better off rejecting a user's data entirely because the privacy-induced noise causes more harm than good to the global model.
  • Truthfulness as a Strategy: Because payoffs are tied to actual accuracy boosts, providing "true" data is a weakly-dominant strategy. Users cannot easily game the system by submitting misleading bids without negatively impacting their own earnings.
Prediction accuracy as a function of privacy levels, highlighting critical points where users should be excluded to maintain model performance.

In the baseline scenario with raw, non-anonymized data, the system achieved a maximum accuracy of 92.5%. As privacy levels (noise variance) increased, prediction accuracy dropped predictably. For example, when a pivotal user increased their privacy beyond a certain point, global accuracy fell from approximately 92% to 90.5%.

Conclusion and Future Work

This research demonstrates that privacy can be effectively integrated into the economic exchange of crowdsensing data. By rewarding marginal contribution, the framework incentivizes high-quality data sharing while respecting user privacy choices. Future work will explore reducing computational costs for calculating Shapley values, implementing fog computing to move processing to the edge, and adapting the model for dynamic environments where users frequently join and leave the network.

Summary updated

Abstract

Mobile crowdsensing has emerged as an efficient sensing paradigm that combines the crowd intelligence and the sensing power of mobile devices, such as mobile phones and Internet of Things gadgets. This article addresses the contradicting incentives of privacy preservation by crowdsensing users, and accuracy maximization and collection of true data by service providers. We first define the individual contributions of crowdsensing users based on the accuracy in data analytics achieved by the service provider from buying their data. We then propose a truthful mechanism for achieving high service accuracy while protecting privacy based on user preferences. The users are incentivized to provide true data by being paid based on their individual contribution to the overall service accuracy. Moreover, we propose a coalition strategy that allows users to cooperate in providing their data under one identity, increasing their anonymity privacy protection, and sharing the resulting payoff. Finally, we outline important open research directions in mobile and people- centric crowdsensing.
BibTeX
@article{alsheikh2017accuracy,
  author = {Mohammad Abu Alsheikh and Yutao Jiao and Dusit Niyato and Ping Wang and Derek Leong and Zhu Han},
  title = {{The Accuracy-Privacy Trade-off of Mobile Crowdsensing}},
  journal = {IEEE Communications Magazine},
  year = {2017},
  doi = {10.1109/mcom.2017.1600737},
}
Figure