Robust and Explainable RIC-Based AI Controllers for 6G RAN Automation

Abstract

AI-driven xApps deployed within the Open Radio Access Network (O-RAN) near-real-time Radio Intelligent Controller (near-RT RIC) are becoming central to mobility optimization, beam management, and load balancing in 5G -Advanced and emerging 6G systems. These controllers operate on tight 10 ms-1 s deadlines, yet machine-learning models are vulnerable to data drift, adversarial perturbations, configuration faults, and unpredictable radio conditions. Unlike deterministic algorithms, ML policies may silently degrade and propagate instability across the RAN. Current O-RAN specifications do not define real-time trust monitoring, certification, or automatic rollback mechanisms for xApps. This paper proposes a lightweight, trust-aware framework that combines KPI-based scoring, behavior-level model analysis, drift and out-of-distribution detection, a compact certification gate, telecom-oriented explainability, and fast rollback logic. A prototype implementation on an OpenAirInterface-based O-RAN testbed demonstrates that the framework detects anomalies within 5-18 seconds, improves handover success rate by 12 percentage points after rollback, and incurs only 3.2% CPU overhead, making it suitable for production deployment.