Abstract
The rapid progression towards sixth-generation (6G) edge-intelligent networks requires Intrusion Detection Systems (IDS) that are interpretable, lightweight, and resilient to adversarial maneuvering. Current deep learning-based IDS systems, including CiNeT, present misleading accuracy through many negative factors. Using the CiNeT framework as the foundation of this paper introduces AARE-IDS, Adaptive Adversarially Resilient Explainable Intrusion Detection System (IDS), for real-time 6G edge settings. AARE-IDS system is particularly exceptional amongst similar developments due to its lightweight bijective traffic encoding (NeT2I-Lite), hybrid CNN - Transformer detector, adversarially trained resilience layer, utilizing a federated learning framework to investigate model aggregation and derived from a differential privacy standard for added security while sharing models. Through extensive experiments on the ToN-IoT, UNSW-NB15, and CICIDS-2018 datasets, AARE-IDS achieved 98.9% accuracy, 39% better robustness against FGSM/PGD attacks, and 45% less CPU overhead compared to CiNeT.

This publication has 12 references indexed in Scilit: