Analysis of malicious input issues on intelligent systems

Intelligent systems can facilitate decision making and have been widely applied to various domains. The output of intelligent systems relies on the users’ input. However, with the development of Web-Based Interface, users can easily provide dishonest input. Therefore, the accuracy of the generated decision will be affected. This dissertation presents three essays to discuss the defense solutions for malicious input into three types of intelligent systems: expert systems, recommender systems, and rating systems. Different methods are proposed in each domain based on the nature of each problem. The first essay addresses the input distortion issue in expert systems. It develops four methods to distinguish liars from truth-tellers, and redesign the expert systems to control the impact of input distortion by liars. Experimental results show that the proposed methods could lead to the better accuracy or the lower misclassification cost. The second essay addresses the shilling attack issue in recommender systems. It proposes an integrated Value-based Neighbor Selection (VNS) approach, which aims to select proper neighbors for recommendation systems that maximize the e-retailer’s profit while protecting the system from shilling attacks. Simulations are conducted to demonstrate the effectiveness of the proposed method. The third essay addresses the rating fraud issue in rating systems. It designs a twophase procedure for rating fraud detection based on the temporal analysis on the rating series.