🧠 Modular Anomaly Detection Framework (MNIST Benchmark)

Overview

This repository implements a modular, end-to-end anomaly detection framework using deep autoencoders, with a focus on reconstruction-based anomaly detection and localization.

The project is designed as a research-ready framework, not a single experiment.
All components—datasets, corruptions, models, metrics, and visualizations—are implemented in a reusable and extensible manner.

The MNIST dataset is used as a controlled benchmark to study:

• What constitutes an anomaly
• How reconstruction error separates normal vs anomalous samples
• How architectural choices (e.g. bottleneck size) affect detection performance
• How anomalies can be localized spatially via error heatmaps

---

Key Contributions

• ✅ End-to-end anomaly detection pipeline
• ✅ Modular corruption (anomaly) generators
• ✅ Autoencoder-based reconstruction model
• ✅ ROC–AUC based evaluation
• ✅ Pixel-level anomaly localization via error heatmaps
• ✅ Robustness studies:
  • Bottleneck size vs performance
  • Contamination robustness
  • Difficulty (noise strength) scaling
• ✅ Script-based reproducibility (no notebook dependency)

---

Repository Structure

.
├── src/
│   ├── __init__.py
│
│   ├── models/
│   │   └── autoencoder.py        # Autoencoder architecture
│
│   ├── datasets/
│   │   └── mnist.py              # MNIST loading utilities
│
│   ├── evaluation/
│   │   └── metrics.py            # Evaluation metrics (ROC, AUC)
│
│   ├── visualization/
│   │   └── heatmaps.py           # Error heatmap & localization utilities
│
│   └── utils/
│       └── corruptions.py        # Anomaly generation functions
│
├── experiments/
│   └── baseline_mnist.py         # End-to-end experiment script
│
├── notebooks/
│   ├── 01_data_exploration.ipynb
│   ├── 02_anomaly_examples.ipynb
│   └── 03_autoencoder_baseline.ipynb
│
├── results/
│   ├── anomalies/                # Heatmaps & localization outputs
│   ├── plots/                    # ROC, robustness, ablation plots
│   └── dataset_stats.json
│
├── requirements.txt
└── README.md

---

Methodology

1. Problem Definition

An anomaly is defined as an input that deviates from the normal data distribution and is therefore poorly reconstructed by a model trained only on normal data.

---

2. Model: Autoencoder

A convolutional autoencoder is trained to reconstruct MNIST digits.

• Encoder compresses the image into a low-dimensional bottleneck
• Decoder reconstructs the image from this representation
• Reconstruction error serves as the anomaly score

---

3. Anomaly Generation

Anomalies are synthetically generated using controlled corruptions:

• Pixel Dropout – random pixel masking
• Gaussian Noise – additive noise with adjustable variance
• Stripe Corruption – full row/column corruption
• Random Patch Occlusion – localized missing regions

These corruptions allow controlled study of anomaly difficulty.

---

4. Detection & Evaluation

• Reconstruction error is computed per image
• ROC curves are generated using normal vs anomalous samples
• Area Under Curve (AUC) is used as the primary metric

---

5. Localization

Pixel-wise reconstruction error maps are visualized as heatmaps to localize anomalous regions within images.

---

6. Robustness Experiments

The framework includes systematic studies of:

• Bottleneck size vs AUC
• Training contamination vs detection performance
• Anomaly difficulty scaling

---

Results Summary

• Near-perfect separation achieved for strong anomalies (AUC ≈ 1.0)
• Performance degrades gracefully with increasing anomaly difficulty
• Smaller bottlenecks encourage better anomaly sensitivity
• Error heatmaps accurately localize corrupted regions

---

How to Run

1. Setup Environment

pip install -r requirements.txt

2. Run Full Experiment

python experiments/baseline_mnist.py

This will automatically:

• Train the autoencoder
• Generate anomalies
• Compute ROC & AUC
• Save plots and heatmaps to results/

---

Design Philosophy

• Modularity over monolithic notebooks
• Reproducibility over ad-hoc experimentation
• Clarity over excessive abstraction
• Research-first, framework-second

---

Future Extensions

This framework is intentionally designed to support:

• Additional datasets (medical imaging, scientific data)
• Alternative models (VAEs, diffusion-based methods)
• Semi-supervised and weakly-supervised anomaly detection
• Benchmarking across domains

---

Author

Ajay Bandiwaddar

---

License

This project is released for research and educational use.

---