Deceptive evidence detection in information fusion of belief functions based on reinforcement learning

Abstract

Counter-deception in information fusion is a significant issue in belief functions, of which how to detect deceptive evidence is a critical problem. We give two definitions of deceptive evidence. Both the feature of the evidence (data) and the feature of the fusion system (combination rule) are considered, and the unsupervised case (without label information) and supervised case (with label information) are also covered. Detecting mathematical model of deceptive evidence is proposed to be solved using the mechanism of reinforcement learning. The proposed reward-shaping method is used to enhance the robustness of the detecting model. For the case without label information, the belief entropy is used to ensure the fusion result of the maintaining evidence to obtain the least uncertainty, and conflict management is utilized to check that the fusion result from each possible state is not counter-intuitive. The compared results of two special examples show the effectiveness of the first definitions of deceptive evidence and the proposed detecting model without label information. In addition, a real application in analyzing the reliability of the feature in DST-based classification is used to illustrate the effect of the second definition of deceptive evidence with label information.