Adversarial Distributions Against Out-of-Distribution Detectors

Sangwoong Yoon, Jinwon Choi, Yonghyeon LEE, Yung-Kyun Noh, Frank C. Park

Keywords: out-of-distribution detection, outlier detection, adversarial attack, model evaluation, markov chain monte carlo

Abstract: Out-of-distribution (OOD) detection is the task of determining whether an input lies outside the training data distribution. As an outlier may deviate from the training distribution in unexpected ways, an ideal OOD detector should be able to detect all types of outliers. However, current evaluation protocols test a detector over OOD datasets that cover only a small fraction of all possible outliers, leading to overly optimistic views of OOD detector performance. In this paper, we propose a novel evaluation framework for OOD detection that tests a detector over a larger, unexplored space of outliers. In our framework, a detector is evaluated with samples from its adversarial distribution, which generates diverse outlier samples that are likely to be misclassified as in-distribution by the detector. Using adversarial distributions, we investigate OOD detectors with reported near-perfect performance on standard benchmarks like CIFAR-10 vs SVHN. Our methods discover a wide range of samples that are obviously outlier but recognized as in-distribution by the detectors, indicating that current state-of-the-art detectors are not as perfect as they seem on existing benchmarks.

One-sentence Summary: We propose a novel evaluation method for out-of-distribution detectors.