Truth Inference in Crowdsourcing Under Adversarial Attacks

Abstract

Crowdsourcing is an information system that provides a cost-effective way of solving computationally challenging problems. However, it is potentially vulnerable to adversarial attacks as the service provider cannot manage workers' behavior. Malicious workers provide unreliable answers to manipulate the system. These attacks affect the truth inference process and thus leads to wrong answers for a targeted set of tasks. Eventually, this reduces the accuracy of aggregated results. Existing works have proposed various types of attacks in crowdsourcing systems and indicate that truth inference is the most affected one. So, we propose methods for defending these attacks for improving the truth inference process. We empirically evaluate the proposed truth inference method on a real and synthetic dataset. The performance of the proposed method is verified, and the results show that it is robust to adversarial attacks with comparable accuracy.