The paper ‘Anomaly Heterogeneity Learning for Open-Set Supervised Anomaly Detection’ proposes a novel approach to anomaly detection, named Anomaly Heterogeneity Learning (AHL). The authors identify a problem with current Open-Set Supervised Anomaly Detection (OSAD) methods, which treat all anomalies as if they come from a single, homogeneous distribution. This limits their ability to detect ‘unseen’ anomalies that may arise from different conditions. AHL addresses this by simulating a variety of anomaly distributions, allowing it to learn a unified model of abnormality. AHL can be integrated with existing OSAD models to improve their performance. Experiments show that it enhances the detection of both seen and unseen anomalies and can generalize to new domains.
Publication date: 20 Oct 2023
Project Page: https://arxiv.org/abs/2310.12790
Paper: https://arxiv.org/pdf/2310.12790
