همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

PE-iForest: A Perturbation-based Framework for Explainable Anomaly Detection

نویسندگان :

Aria Shakoo¹ Mahboobeh Riahi-Madvar²

1- دانشگاه شهید بهشتی 2- دانشگاه شهید بهشتی

کلمات کلیدی :

anomaly detection،explainable AI (XAI)،Isolation Forest،perturbation explanation،SHAP،DIFFI،scalability

چکیده :

As anomaly detection is increasingly deployed in high-stakes domains such as healthcare, finance, and cybersecurity, there is a critical need for methods that are simultaneously interpretable, scalable, and accurate. Explainable anomaly detection methods often suffer from high computational complexity in model-agnostic approaches or are limited by reliance on a single model in model-specific methods. This paper introduces a perturbation-based framework for anomaly detection that leverages vectorized perturbations to deliver high-speed explanations while remaining agnostic to the underlying detection algorithm. We instantiate the framework on Isolation Forest (iForest), a widely used scalable detector, to address its inherently “blackbox” nature and enhance user trust through interpretable outputs. We empirically evaluate PE-iForest against state-of-the-art explainers DIFFI, TreeSHAP and traditional detection baselines (LOF, HBOS, PCA) using four public benchmark datasets. Our optimized implementation achieves superior efficiency, generating explanations for 100 anomalies in 0.89 seconds on highdimensional Arrhythmia data, compared to 2.87 seconds for DIFFI. PE-iForest also yields explanations that are qualitatively and quantitatively aligned with more complex methods. Overall, the results of PE-iForest demonstrate that our simple, vectorized perturbation framework can outperform complex methods in speed while maintaining high trustworthiness.