FMMO: Detecting the Divergence Between Local Attribution and Global Drift

Local explanation methods can mask serious problems: Popular explainability tools like TreeSHAP can appear stable and trustworthy even when a model's actual performance is collapsing after deployment, creating a false sense of security for organizations relying on these tools for auditing.

Hidden fairness failures go undetected: The paper demonstrates that local attribution methods fail to flag when model drift causes disparate impact on protected groups—specifically showing cases where False Positive Rates spike for certain demographics while feature explanations remain unchanged, meaning discrimination can occur invisibly.

Global monitoring catches what local explanations miss: By combining global surrogate models with usage measurements, FMMO successfully identifies discriminatory deterioration that standard local XAI tools overlook, suggesting organizations need multiple monitoring approaches rather than relying on a single explanation method.

Delayed labels create accountability gaps: The paper highlights that when ground truth labels arrive late in real-world deployments, "silent failures" can persist undetected—a particular risk for high-stakes decisions where fairness violations may harm people before anyone notices performance has degraded.

Explainability alone is insufficient for responsible AI: This research shows that explainability and fairness monitoring are distinct challenges; having interpretable models doesn't guarantee you'll catch when those models start making biased decisions in production.