Managing trust in diffusion adaptive networks with malicious agents

In this paper, we consider the problem of information sharing over adaptive networks, where a diffusion strategy is used to estimate a common parameter. We introduce a new model that takes into account the presence of both selfish and malicious intelligent agents that adjust their behavior to maximize their own benefits. The interactions among agents are modeled as a stochastic game with incomplete information and partially observable actions. To stimulate cooperation amongst selfish agents and thwart malicious behavior, a trust management system relying on a voting scheme is employed. Agents act as independent learners, using the Q-learning algorithm. The simulation results illustrate the severe impact of falsified information on estimation accuracy along with the noticeable improvements gained by stimulating cooperation and truth-telling, with the proposed trust management mechanism.