XAI-on-RAN: Explainable, AI-native, and GPU-Accelerated RAN Towards 6G

TL;DR

The paper tackles the challenge of deploying AI-native RAN control in 6G for mission-critical verticals by demanding transparency and accountability in AI decisions. It introduces XAI-on-RAN with an AI-native GPU-accelerated architecture and a novel xAI-Native xApp that provides real-time explanations using both intrinsic Attention and post-hoc Integrated Gradients, while integrating a framework to analyze fidelity and latency trade-offs through metrics like $\Phi$ and local $R^{2}$, and the latency decomposition $T_{total} = T_{inf} + T_{xai} + T_{comm}$. Empirical results show that the hybrid Attention+IG approach achieves the best fidelity-latency balance in online operation, whereas SHAP is more suitable for offline auditing and Attention alone is fast but less faithful. The work demonstrates the practical viability of deploying explainable AI within GPU-accelerated RANs, enabling trustworthy, fair, and reliable AI-assisted network management for URLLC and vertical applications in emerging 6G ecosystems, while highlighting design trade-offs for real-time interpretability.

Abstract

Artificial intelligence (AI)-native radio access networks (RANs) will serve vertical industries with stringent requirements: smart grids, autonomous vehicles, remote healthcare, industrial automation, etc. To achieve these requirements, modern 5G/6G design increasingly leverage AI for network optimization, but the opacity of AI decisions poses risks in mission-critical domains. These use cases are often delivered via non-public networks (NPNs) or dedicated network slices, where reliability and safety are vital. In this paper, we motivate the need for transparent and trustworthy AI in high-stakes communications (e.g., healthcare, industrial automation, and robotics) by drawing on 3rd generation partnership project (3GPP)'s vision for non-public networks. We design a mathematical framework to model the trade-offs between transparency (explanation fidelity and fairness), latency, and graphics processing unit (GPU) utilization in deploying explainable AI (XAI) models. Empirical evaluations demonstrate that our proposed hybrid XAI model xAI-Native, consistently surpasses conventional baseline models in performance.

Figures (2)

• Figure 1: End-to-end XAI-Native testbed with GPU-accelerated RAN.
• Figure 2: Comparative evaluation with baseline models.