Madsgm: Multivariate Anomaly Detection with Score-Based Generative Models

By Haksoo Lim et al.

Published on Oct. 21, 2023

Read the original document by opening this link in a new tab.

Table of Contents

ABSTRACT
KEYWORDS
INTRODUCTION
RELATED WORKS
Anomaly Detection
Score-Based Generative Models
Adversarial Purification
PROPOSED METHOD
Problem Statement
Score Network for Time-series Anomaly Detection
Training and Sampling Methods
Anomaly Measurement
Adversarial Purification
EXPERIMENTS
Benchmark Datasets
Evaluation Metrics
Results
CONCLUSION
Future Work
REFERENCES

Summary

The document presents a multivariate time-series anomaly detector named Madsgm, based on score-based generative models. It addresses the challenges in time-series anomaly detection, emphasizing the need for unsupervised training due to the difficulty in labeling anomalous observations. Madsgm considers three types of anomaly measurements - reconstruction-based, density-based, and gradient-based, to provide robust detection. The proposed method utilizes a conditional score network and denoising score matching loss for training. Experiments on benchmark datasets demonstrate the effectiveness of Madsgm in achieving accurate predictions compared to baselines. The contributions of the study include the introduction of SGMs for time-series anomaly detection, training a conditional score network for this purpose, and conducting comprehensive experiments to validate the model's performance.