Explanation-Guided Fair Federated Learning for Transparent 6G RAN Slicing

Explainability should be a measurable service requirement: The paper proposes that AI faithfulness—how well an AI system's explanations actually reflect its decision-making—should be treated as a formal service-level agreement (SLA) metric in 6G networks, alongside traditional performance measures. This means organizations should contractually commit to explaining their AI systems, not just achieving accuracy.

Fairness constraints improve real-world outcomes: By incorporating fairness goals (measured through recall) directly into the AI training process, the proposed method improved fairness performance by over 25% compared to unconstrained approaches. This demonstrates that fairness isn't just an ethical add-on but can be systematically built into AI systems from the ground up.

Explanation quality can be quantitatively measured and improved: The research shows that explanation quality (measured by "comprehensiveness score") can be objectively assessed and improved during training—achieving over 50% improvement compared to baseline methods. This suggests organizations can move beyond subjective judgments about whether their AI explanations are trustworthy.

Federated learning enables trustworthy AI without centralizing data: By using federated learning (where AI training happens across distributed networks rather than in one location), the approach maintains privacy while still building transparent, fair AI systems—addressing concerns that explainability and fairness require exposing sensitive data.

Technical design choices matter for responsible AI: The paper shows that specific mathematical techniques (Jensen-Shannon divergence versus Kullback-Leibler divergence) significantly impact both explanation quality and fairness outcomes, suggesting that responsible AI requires careful technical implementation, not just good intentions.