Abstract:
Abstract: Detecting out-of-distribution (OOD) data is essential for model reliability. I present a convex-hull (CH)–based anomaly detection method that exploits the observation that OOD samples expand the CH volume. By iteratively removing points and tracking volume changes, the method establishes a decision boundary between in- and out-of-distribution data. Tested against seven baselines on ten datasets, it achieves performance comparable to state-of-the-art techniques and includes an efficient criterion for identifying when this approach offers the greatest advantage.
