Joint Explainability and Sensitivity-Aware Federated Deep Learning for Transparent 6G RAN Slicing

TL;DR

This work tackles trustworthy, zero-touch management of 6G network slices by introducing a closed-loop Explainable Federated Deep Learning (FDL) framework for traffic drop classification in RAN. The approach jointly optimizes predictive accuracy with recall guarantees and explainability fidelity, using a log-odds attribution metric as a constraint within a proxy-Lagrangian two-player game, and leverages Integrated Gradients to produce feature attributions fed back into the learning loop. Compared to a vanilla, post-hoc Integrated Gradient baseline, the proposed method achieves faster convergence and higher trustworthiness, evidenced by improved recall, favorable log-odds trajectories, and attribution heatmaps that reveal the most impactful features, such as channel quality. The methodology enables non-IID data across distributed base stations while preserving transparency and SLA-aligned performance, advancing practical, explainable AI for 6G network slicing.

Abstract

In recent years, wireless networks are evolving complex, which upsurges the use of zero-touch artificial intelligence (AI)-driven network automation within the telecommunication industry. In particular, network slicing, the most promising technology beyond 5G, would embrace AI models to manage the complex communication network. Besides, it is also essential to build the trustworthiness of the AI black boxes in actual deployment when AI makes complex resource management and anomaly detection. Inspired by closed-loop automation and Explainable Artificial intelligence (XAI), we design an Explainable Federated deep learning (FDL) model to predict per-slice RAN dropped traffic probability while jointly considering the sensitivity and explainability-aware metrics as constraints in such non-IID setup. In precise, we quantitatively validate the faithfulness of the explanations via the so-called attribution-based \emph{log-odds metric} that is included as a constraint in the run-time FL optimization task. Simulation results confirm its superiority over an unconstrained integrated-gradient (IG) \emph{post-hoc} FDL baseline.

Figures (6)

• Figure 1: RAN federated traffic drop classification in NS
• Figure 2: Explainable FDL building blocks
• Figure 3: Analysis of FL training loss vs FL rounds of Proposed EFL with Lower bound of Recall score, $\alpha = [0.9, 0.95, 0.95]$ and Upper bound of log-odds score, $\beta = [-0.01, -0.01 , -0.01]$
• Figure 4: Analysis of Recall score with Lower bound of Recall score, $\alpha = [0.9, 0.95, 0.95]$ and Upper bound of log-odds score, $\beta = [-0.01, -0.01 , -0.01]$
• Figure 5: Analysis of log-odds score with Lower bound of Recall score, $\alpha = [0.9, 0.95, 0.95]$ and Upper bound of log-odds score, $\beta = [-0.01, -0.01 , -0.01]$