Generative Perturbation Analysis for Probabilistic Black-Box Anomaly Attribution

Standard explainability methods have limitations for anomaly detection: The paper identifies that popular XAI techniques like Shapley values are "deviation-agnostic," meaning they cannot effectively explain why a specific prediction is anomalous rather than explaining how the model generally works. This suggests that organizations using standard explainability tools for anomaly detection may be getting incomplete information about what went wrong.

Uncertainty quantification matters for anomaly attribution: Rather than providing a single attribution score for each variable, this framework produces a probability distribution showing the range of possible explanations. This allows practitioners to understand not just which variables caused an anomaly, but how confident they should be in that attribution—critical for high-stakes decisions.

Black-box anomaly explanation is now feasible without training data: The method works without access to the original training dataset, which is practically important since organizations often cannot share sensitive training data with external auditors or stakeholders investigating anomalies. This makes responsible AI auditing more feasible in real-world scenarios.

The framework uses counterfactual reasoning: By modeling how input variables would need to change to make an anomalous observation "normal" again, the approach provides intuitive, actionable explanations that stakeholders can understand and act upon—moving beyond abstract statistical measures.