DLFTI: A deep learning based fast truth inference mechanism for distributed spatiotemporal data in mobile crowd sensing

Abstract

The paradigm of Mobile Crowd Sensing (MCS) allows for numerous applications with distributed spatiotemporal data, where great attention is drawn to the fundamental problems for truth inference. Existing works all suffer from poor accuracy and slow start, due to the lack of valid information for Ground Truth Data (GTD) and workers. These problems make the MCS system vulnerable to fraudulent attacks by malicious gangs, causing the Estimated Truth Data (ETD) to deviate significantly from GTD. In this paper, we propose a Deep Learning based Fast Truth Inference mechanism, called DLFTI, to achieve fast trust computing and accurate truth discovery in MCS. First, we introduce the Degrees-Of-Trust (DOT) to characterize the sensing ability of workers and establish worker profiles based on DOT to recognize workers’ trustworthiness dynamically. Then, we abandon the unrealistic assumption of priori GTD in previous studies and instead utilize the Unmanned Aerial Vehicles (UAVs), recognized trustworthy workers and the Deep Matrix Factorization (DMF) method to construct three-level GTD and three-level ETD, which are used for fast trust computing of workers and accurate truth discovery of tasks respectively. Finally, we conduct extensive simulations on a real-world dataset to corroborate the significant performance of DLFTI.