Variational Autoencoder for Anomaly Detection

Probabilistic anomaly detection using a Variational Autoencoder (VAE). Learns the latent distribution of normal data during training, then detects anomalies by computing reconstruction probability at inference time. Provides calibrated uncertainty quantification alongside anomaly scores through Monte Carlo sampling in the latent space.

Architecture

Input x
   |
   v
Encoder: FC(input_dim, 128) -> ReLU -> FC(128, 64) -> ReLU
   |                                       |
   v                                       v
FC(64, latent_dim) = mu          FC(64, latent_dim) = log_var
   |                                       |
   +---> Reparameterization Trick <--------+
   |     z = mu + eps * exp(0.5 * log_var)
   v
Decoder: FC(latent_dim, 64) -> ReLU -> FC(64, 128) -> ReLU -> FC(128, input_dim) -> Sigmoid
   |
   v
Reconstruction x_hat

Loss = Reconstruction Loss (BCE/MSE) + beta * KL Divergence

Comparison with Autoencoder + Isolation Forest (Primary Approach)

Feature	VAE (this project)	AE + Isolation Forest (anomaly-detection-system)
Anomaly Score	Reconstruction probability	Reconstruction error + isolation score
Uncertainty	Native (posterior sampling)	Requires calibration
Interpretability	Moderate (latent space)	High (feature importance)
Training Stability	Requires KL annealing	Straightforward
Threshold Setting	Probability-based	Percentile-based (simpler)
Robustness	KL collapse risk	More robust

Decision: The simpler Autoencoder + Isolation Forest ensemble was chosen for the primary project due to better interpretability, more stable training, and fewer hyperparameters to tune. The VAE approach provides richer uncertainty estimates but requires careful KL annealing and monitoring to avoid posterior collapse.

Setup

Prerequisites

• Python 3.10+
• PyTorch 2.0+

Installation

python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Running

# Start API server
uvicorn src.api:app --host 0.0.0.0 --port 8000 --reload

# Generate demo data and train
python -m src.data_generator
python -m src.trainer --config configs/config.yaml

API Endpoints

Method	Endpoint	Description
POST	/detect	Anomaly detection with uncertainty estimates
POST	/encode	Get latent space representation
POST	/train	Train a new model on uploaded data
GET	/health	Service health check

Usage Example

import requests
import numpy as np

# Detect anomalies
data = np.random.randn(10, 30).tolist()
response = requests.post(
    "http://localhost:8000/detect",
    json={"data": data}
)
results = response.json()
for i, r in enumerate(results["results"]):
    print(f"Sample {i}: anomaly={r['is_anomaly']}, score={r['anomaly_score']:.4f}")

Project Structure

vae-anomaly-detection/
├── src/
│   ├── vae_model.py          # VAE architecture
│   ├── anomaly_detector.py   # Detection pipeline
│   ├── trainer.py            # Training with KL annealing
│   ├── data_generator.py     # Synthetic data generation
│   └── api.py                # FastAPI application
├── configs/
│   └── config.yaml           # Configuration
├── requirements.txt
└── README.md

License

MIT License