Trust-Based Selective Collaboration for Multi-Agent Intrusion Detection in IIoT

In distributed intrusion detection systems for the Industrial Internet of Things (IIoT), collaboration among heterogeneous agents can compensate for each agent’s partial competence. However, when done indiscriminately, it exposes the agents most capable of negative transfer, degrading decisions that would otherwise be correct on their own. This work reframes the problem as an admission problem, shifting the question from whether to collaborate to with whom to collaborate, and proposes a mechanism in which each agent maintains, in a strictly local manner, a trust matrix that directs its requests for help only to peers whose historical contributions have proven beneficial, without altering the system’s aggregation rule. The study compares four trust-updating strategies (exponential moving average, Beta-Bernoulli, upper confidence bound, and binary) and two admission policies (threshold and top-k) against baseline strategies of isolated decision-making and indiscriminate collaboration, on a testbed with ten heterogeneous agents running on real hardware (Raspberry Pi) and using two benchmark datasets with different degrees of heterogeneity (UNSW-NB15 and CICIoT2023). The results show that selective collaboration based on trust preserves the gains of weak agents and suppresses the degradation of strong agents. The benefit is moderate under purely algorithmic heterogeneity, but becomes the factor that makes collaboration worthwhile when agents are trained on distinct datasets, where naive collaboration offers almost no benefit. The trust mechanism also reduces the total message volume by 44% to 84% depending on the configuration (exceeding 60% under top-k admission) and cuts collaboration latency by approximately 44% to 77% depending on the configuration, without compromising resources, and with its value growing with the degree of heterogeneity, from marginal when agents share knowledge to the factor that decides whether collaboration is beneficial when they do not share it.