Scalable, Explainable and Provably Robust Anomaly Detection with One-Step Flow Matching

Authors: Zhong Li, Qi Huang, Yuxuan Zhu, Lincen Yang, Mohammad Mohammadi Amiri, Niki van Stein, Matthijs van Leeuwen.

Time-Conditioned Contraction Matching (TCCM) adapts ideas from flow matching for semi-supervised anomaly detection in tabular data. It scores anomalies in a single forward pass, supports feature-level explanations, and provides robustness guarantees through a Lipschitz-continuous score. Evaluations on ADBench compare its detection accuracy and computational cost with existing methods.