A privacy-protected intelligent crowdsourcing application of IoT based on the reinforcement learning

The crowdsourcing scheme emerges as a promising solution for data-based application in the Internet of Things (IoT) network by dividing the large-scale complex sensing tasks into simple micro sensing tasks. AI technologies have been widely applied in crowdsourcing applications for IoT security. However, there are still several issues to be stressed. There exist malicious participants aiming at gaining unjust payment. A trust evaluation mechanism can effectively filter the attackers. Nevertheless, the existing trust evaluation mechanisms cannot preclude co-cheating and overlook the conflicts with the privacy exposure of participants. A novel Privacy-protected Intelligent Crowdsourcing scheme based on Reinforcement Learning (PICRL) is proposed. PICRL optimizes the utility of the system considering the data amount, data quality, and costs at the same time. The main innovations of the PICRL are as follows. First, the quality is guaranteed by an effective trust evaluation mechanism. The proposed trust evaluation consists of three parts: privacy trust, crowd trust, and hybrid active trust. Second, the trust evaluation can effectively prevent co-cheating and provide personal privacy exposure choice for the participants. Third, PICRL maximizes the utility based on evaluated trust utilizing the reinforcement method Q-learning without knowing the specific sensing model, which aims at maximizing the cumulated reward by the selection of state-action pair. The effectiveness of the proposed PICRL is verified by the extensive simulation experiments. (C) 2021 Elsevier B.V. All rights reserved.

Automated summary

The crowdsourcing scheme for data-based applications in the IoT network faces issues with malicious participants, which a proposed privacy-protected intelligent crowdsourcing scheme (PICRL) aims to address using reinforcement learning and trust evaluation mechanisms. The PICRL optimizes utility through effective trust assessment and Q-learning method without prior knowledge of specific sensing models.