Evaluation of Applying Federated Learning to Distributed Intrusion Detection Systems Through Explainable AI

We apply federated learning (FL) to a distributed intrusion detection system (IDS), in which we deploy numerous detection servers on the edge of a network. FL can mitigate the impact of decreased training data in each server and exhibit almost the same detection rate as that of the non-distributed IDS for all attack classes. We verify the effect of FL using explainable artificial intelligence (XAI); this effect is demonstrated by the distance between the feature set of each attack class in the distributed IDS and that in the non-distributed IDS. The distance increases for independent learning and decreases for FL.