Stable Diffusion for Anomaly Detection in Multivariate Time Series

Overview

This project implements a Stable Diffusion Model for anomaly detection in multivariate time series data. The model leverages anomaly scores computed using the MTAD-GAT model as a conditioning input in the reverse diffusion process to improve anomaly detection accuracy.

Objective

The model predicts whether a given timestamp is anomalous or normal, making it suitable for real-world anomaly detection tasks in time-series data.

Key Features

✅ Stable Diffusion Model for anomaly detection
✅ Anomaly scores from MTAD-GAT used as conditioning input
✅ Reverse diffusion process incorporates anomaly scores
✅ Best F1-score optimization achieved: 0.96 F1-score
✅ Requires a GPU for training on a new dataset

Performance

• F1-Score: 0.96 (using best F1-score search method)

Applications

This model can be applied to various time-series anomaly detection tasks, including:

• 🏦 Fraud Detection
• 🏭 Industrial Equipment Monitoring
• 🌐 Network Intrusion Detection
• 🏥 Healthcare Anomaly Detection

Installation & Usage

1️⃣ Setup AWS EC2 Instance (GPU Recommended)

• Launch an AWS EC2 instance with a **GPU

2️⃣ Clone the repository

git clone https://github.com/JisnaP/stableDiffTSAD.git
cd stableDiffTSAD