Explanation-Based Runtime Verification for Trustworthy ML-driven Optical Networks

ML decisions in optical networks have immediate real-world consequences: The paper highlights that ML predictions in network automation are "directly coupled with control-plane actions where incorrect decisions can immediately impact service quality, resource efficiency, and network stability," making reliability verification essential before deployment.

Explainability can serve a verification function beyond transparency: Rather than using explanations solely to understand why a model made a decision, this work proposes using them as a runtime safety mechanism to check whether individual predictions are trustworthy before they control actual network operations.

Two specific verification checks can catch problematic decisions: The approach evaluates "explanation coherence" (whether the explanation makes logical sense) and "physics grounding consistency" (whether the explanation aligns with known physical laws of optical networks), allowing the system to defer or reject uncertain decisions.

The method enables graceful degradation under uncertainty: By flagging questionable ML decisions at runtime rather than blindly executing them, the system can "defer or reject decisions" rather than allowing potentially harmful automated actions, providing a safety valve for high-stakes applications.

This addresses a critical gap in ML deployment for critical infrastructure: The paper tackles the challenge of ensuring "reliability of individual decisions at deployment time" in systems where automation is increasing, suggesting that explanation-based verification could be particularly valuable for other critical domains beyond optical networks.