[Paper Review] Towards Bridging the FL Performance-Explainability Trade-Off: A Trustworthy 6G RAN Slicing Use-Case

Swastika Roy, Hatim Chergui+1 more|arXiv (Cornell University)|Jul 24, 2023

Privacy-Preserving Technologies in Data4 citations

TL;DR

This paper proposes an explanation-guided in-hoc federated learning (FL) framework that integrates explainable AI (XAI) into FL training to simultaneously optimize performance and explainability in 6G RAN slicing. By using a closed-loop feedback between a resource allocation model and an explainer—guided by a confidence metric based on feature attributions (IG, SHAP, Input×Gradient)—the method achieves superior model trustworthiness and faster convergence than post-hoc FL baselines, especially under non-IID data.

ABSTRACT

In the context of sixth-generation (6G) networks, where diverse network slices coexist, the adoption of AI-driven zero-touch management and orchestration (MANO) becomes crucial. However, ensuring the trustworthiness of AI black-boxes in real deployments is challenging. Explainable AI (XAI) tools can play a vital role in establishing transparency among the stakeholders in the slicing ecosystem. But there is a trade-off between AI performance and explainability, posing a dilemma for trustworthy 6G network slicing because the stakeholders require both highly performing AI models for efficient resource allocation and explainable decision-making to ensure fairness, accountability, and compliance. To balance this trade off and inspired by the closed loop automation and XAI methodologies, this paper presents a novel explanation-guided in-hoc federated learning (FL) approach where a constrained resource allocation model and an explainer exchange -- in a closed loop (CL) fashion -- soft attributions of the features as well as inference predictions to achieve a transparent 6G network slicing resource management in a RAN-Edge setup under non-independent identically distributed (non-IID) datasets. In particular, we quantitatively validate the faithfulness of the explanations via the so-called attribution-based confidence metric that is included as a constraint to guide the overall training process in the run-time FL optimization task. In this respect, Integrated-Gradient (IG) as well as Input $ imes$ Gradient and SHAP are used to generate the attributions for our proposed in-hoc scheme, wherefore simulation results under different methods confirm its success in tackling the performance-explainability trade-off and its superiority over the unconstrained Integrated-Gradient post-hoc FL baseline.

Motivation & Objective

• Address the performance-explainability trade-off in AI-driven 6G network slicing, where high model accuracy and transparency are both required for trustworthy zero-touch management.
• Overcome the limitations of post-hoc XAI methods that lack integration with model training, leading to unreliable or non-optimized explanations.
• Enable transparent, accountable, and compliant AI decision-making in distributed RAN-Edge environments with non-IID data distributions.
• Introduce a closed-loop mechanism where model predictions and feature attributions are iteratively refined to improve both performance and explanation fidelity.
• Quantify explanation faithfulness using a novel attribution-based confidence metric embedded in the FL optimization process.

Proposed method

• Propose a closed-loop (CL) in-hoc FL framework where a constrained resource allocation model and an XAI explainer exchange soft attributions and predictions iteratively.
• Integrate a proxy-Lagrangian two-player game strategy to jointly optimize model performance and explanation faithfulness via a confidence metric.
• Use Integrated Gradients (IG), Input×Gradient, and SHAP as attribution generators to produce feature importance scores during training.
• Embed the attribution-based confidence metric as a constraint in the run-time FL optimization to ensure faithful and reliable explanations.
• Implement the in-hoc FL scheme in a RAN-Edge setup with non-IID data, simulating real-world 6G network slicing scenarios.
• Evaluate convergence time and model performance across different XAI methods and compare against an unconstrained post-hoc FL baseline.

Figure 2: in-hoc FL for transparent resource allocation.

Experimental results

Research questions

• RQ1 Can an in-hoc FL approach that integrates XAI during training effectively reduce the performance-explainability trade-off in 6G network slicing?
• RQ2 How does the integration of a confidence metric based on feature attributions affect the fidelity and reliability of explanations in FL?
• RQ3 What is the impact of different XAI methods (IG, SHAP, Input×Gradient) on convergence speed and model confidence in the in-hoc FL framework?
• RQ4 How does the proposed in-hoc FL method compare to a post-hoc FL baseline in terms of training time and explanation quality under non-IID data?
• RQ5 Which network parameters (e.g., CQI, OTT traffic, MIMO full rank) have the most significant influence on CPU allocation predictions, and can they be reliably identified through the proposed method?

Key findings

• The in-hoc FL approach with Integrated Gradients (IG) as the attribution method achieves the fastest convergence time—close to the post-hoc baseline—while maintaining the highest prediction confidence.
• The attribution-based confidence metric successfully quantifies explanation faithfulness and acts as an effective constraint in the FL optimization process.
• SHAP-based attributions reveal that CQI has the strongest positive impact on CPU allocation, indicating that better channel quality leads to reduced CPU load due to fewer retransmissions and lower queuing delays.
• OTT traffic per TRP shows a concentration of positive values, indicating that higher traffic levels correlate with increased CPU load, enabling dynamic resource scaling.
• MIMO full rank has a negligible impact on CPU load in this scenario due to its low magnitude, though the model can still detect potential signal degradation issues.
• The in-hoc FL framework demonstrates lower computational complexity up to convergence than the post-hoc baseline, while delivering superior explanation confidence and performance.

Figure 3: FL training MSE loss vs. FL rounds with $\nu=[0.82,0.83,0.85]$