Soteria: A Privacy-Aware and Budget-Limited Multi-Armed Bandit-Based Crowdsourcing Approach for Edge Video Analytics
Zhihua Wang, Yuan Ji, Jieying Zhou, Minghui Jin, Mengjiang Zhu · 2024
In recent years, crowdsourcing has become a cost-effective approach to video analytics due to the large number of workers using mobile devices with image processing capabilities based on computer vision technologies. Crowdsourcing workers are hired to perform video analytics tasks and then rewarded based on the quality of their work. However, the crowdsourcing platform may face several challenges when assigning video analytics tasks to appropriate workers: jointly optimizing analytics accuracy and energy-related resource utilization is challenging due to the inherent uncertainty in accuracy and constraints imposed by workers’ configurations; workers’ analytics accuracy reflects their privacy information, including their devices’ computing capabilities, which must be protected. Considering those challenges for video analytics crowdsourcing, we first formulate an online optimization problem that maximizes profits based on accuracy within a limited budget of energy-related resource. To tackle it, we design a budget-limited multi-armed bandit-based algorithm Soteria for video dispatch, which captures the changes in analytics accuracy and then exploits them by efficiently solving knapsack problems. Besides, a δ-differential privacy mechanism is introduced to guarantee workers’ privacy. We prove Soteria’s effectiveness in setting a regret bound, and extensive experiments show the superior performance of Soteria compared to others in terms of crowdsourcing utilities by up to 9.7%.