Leveraging Heterogeneous Side Information via Diffusion Models for Time-series Anomaly Detection

Hanchen Xiong, Janmey Shukla, Aakarshan Chauhan, Zicun Cong, Dianhuan Lin

Keywords: diffusion models, time series, anomaly detection

Abstract: In this paper, we propose a novel unsupervised learning approach for Out-of- Distribution (OoD) detection in time-series data, leveraging state-of-the-art dif- fusion models to capture the underlying data distribution. Our first contribution is the development of an effective OoD detector based on conditional sampling and reconstruction error measurement, eliminating the need for labeled data samples. We employ time series imputation techniques to reconstruct the original time se- ries, enhancing the detection process. Our second contribution is the incorporation of domain-specific side information, which bolsters the diffusion model’s ability to understand the structure of time-series data and results in a more robust OoD detector. Finally, we demonstrate the state-of-the-art performance of our proposed method through experiments on three diverse datasets: IoT Event Sequence De- tection, DDoS Attack Detection, and Abnormal Network Transaction Sequence Detection. The experimental results highlight the effectiveness and versatility of our approach in addressing various OoD detection tasks across different domains

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

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/2024/AuthorGuide.

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: 5763