A Q-Learning Based Multi-Agent Mechanism for Collusion Tolerance in Voting-Based Systems

Many critical systems use voting-based mechanisms to mask or tolerate faults and improve system reliability. However, such systems are susceptible to collusion attacks where collaborating malicious resources may produce identical erroneous results, potentially causing the task to fail. To detect and alleviate the consequence of such coordinated adversarial behavior, we develop a multi-agent collusion tolerance system, systematically integrating Q-learning based agent allocation, dual-stage spot checking, and strategic credibility adjustment. A spotter agent is in charge of detecting a malicious resource agent based on its performance at the pre-task execution stage and at the post-voting stage. In the case of collusion being detected, the credibility scores of collusive participants are revoked, excluding them from future allocations and task execution. To balance resource utilization and defense effectiveness, optimally allocating spotter agents and resource agents is a crucial and challenging problem. We address this challenge by using reinforcement learning, particularly, Q-learning to optimally allocate these agents under a predefined cost constraint. Comprehensive experiments are performed to illustrate the effectiveness of the proposed collusion tolerance mechanism. A sensitivity analysis of key model parameters is also conducted to demonstrate their impact on the performance of the proposed mechanism.