Narcissus: Leveraging Early Training Dynamics for Unsupervised Anomaly Detection

Qiyu Wang, Chuanhao Sun, Brooks Paige, Kai Xu, Mahesh K. Marina

Keywords: Anomaly Detection, Unsupervised Learning, Early Stopping

TL;DR: We realize an unsupervised training scheme to use semi-supervised anomaly detection method without label without loss on accuracy.

Abstract: Anomaly detection is a critical learning task with many significant and diverse applications. Currently, semi-supervised methods provide the state-of-the-art accuracy performance but require labeled normal data for training. Unsupervised approaches, on the other hand, do not have this requirement but can only offer inferior anomaly detection performance. In this paper, we introduce NARCISSUS, a novel unsupervised anomaly detection method that achieves accuracy comparable to semi-supervised approaches. Our key insight is that a learning model when training with a mix of normal and sparse anomalous data converges first on normal data. Leveraging this insight, NARCISSUS employs a tailored early stopping scheme, eliminating the need for pseudo labels and costly label generation interactions. It also offers systematic solutions to minimize the influence of model uncertainty, ensuring robust detection. NARCISSUS is model-agnostic and can therefore make use of even a semi-supervised anomaly detection model underneath, thereby turning it into an unsupervised one. Comprehensive evaluations using time series, image and graph datasets show that NARCISSUS provides similar or better detection performance compared to best-performing semi-supervised methods while not requiring labeled data.

Primary Area: learning on time series and dynamical systems

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 9839