A Survey of Unsupervised Learning-Based Out-of-Distribution Detection

Out-of-distribution (OOD) detection is the task of distinguishing abnormal data that lies outside the bounds of the training dataset's distribution. OOD detection plays a vital role in various applications of machine learning and deep learning, including intrusion detection in cybersecurity, diagnostics in med-ical data, and defect classification in manufacturing processes. While models for OOD detection are typically trained using supervised learning, this approach requires significant cost and effort such as collection and labeling of OOD data. To address this issue, unsupervised learning-based methods have been proposed, which can overcome the drawbacks of supervised approaches. In this paper, we introduce generative model-based OOD methods and self-supervised OOD detection methods within the realm of unsupervised learning. We also analyze the performance of state-of-the-art unsupervised learning-based OOD methods to suggest future research directions.