# Time Series Anomaly Detection using Diffusion-based Models

**Authors:** Ioana Pintilie, Andrei Manolache, Florin Brad  
**arXiv:** [2311.01452](https://arxiv.org/abs/2311.01452)  
**Date:** 2023-11-02  
**Summary page:** https://www.summarizepaper.com/en/paper/2311.01452v1/

## Key points

- Diffusion-based models are explored for anomaly detection in multivariate time series
- Two diffusion-based models are tested and compared to strong neural baselines
- An extension of the PA%K protocol called ROCK-AUC is introduced, which is agnostic to detection threshold and ratio K of correctly detected points
- Experimental results show that both diffusion-based models outperform the autoencoder model based on F 1K-AUC score
- DiffusionAE model performs better than Diffusion model by 4 F 1K-AUC points and exhibits smaller variance
- DiffusionAE model is more robust to different anomaly types in datasets and suitable for realistic scenarios with anomalous events producing abnormal values on multiple dimensions simultaneously
- Diffusion process smooths out anomaly segments, leading to higher reconstruction errors and improved detection
- ROCK-AUC can be used as a metric for model selection when evaluating diffusion-based anomaly detection models
- DiffusionAE model achieves better overall performance and greater robustness to different training anomaly ratios and multiple anomaly types within a dataset
- Findings highlight the potential of diffusion-based methods for AD in MTS.

## Summary

This paper explores the use of diffusion-based models for anomaly detection (AD) in multivariate time series (MTS). The authors test two diffusion-based models and compare them to several strong neural baselines. They introduce an extension of the PA%K protocol called ROCK-AUC, which is agnostic to both the detection threshold and the ratio K of correctly detected points. The experimental results show that both diffusion-based models outperform the autoencoder model based on the F 1K-AUC score. Specifically, the DiffusionAE model performs better than the Diffusion model by 4 F 1K-AUC points and exhibits smaller variance. This indicates that the DiffusionAE model is more robust to different anomaly types in datasets and may be more suitable for realistic scenarios where anomalous events produce abnormal values on multiple dimensions simultaneously. The diffusion process smooths out anomaly segments, leading to higher reconstruction errors and improved detection. Additionally, ROCK-AUC can be used as a metric for model selection when evaluating diffusion-based anomaly detection models. The DiffusionAE model achieves better overall performance and greater robustness to different training anomaly ratios and multiple anomaly types within a dataset. Overall, these findings highlight the potential of diffusion-based methods for AD in MTS.