Autoencoder Optimization for Anomaly Detection

This repository holds the notebook and supporting code behind our paper, "Autoencoder Optimization for Anomaly Detection: A Comparative Study with Shallow Algorithms", presented at IEEE IJCNN 2024.

Live Demo

huggingface.co/spaces/VisheshSrivastava/autoencoder-anomaly-detection — upload an image and see the trained autoencoder flag it normal or anomalous, with the reconstruction error map alongside it.

What's here

We ask how far a plain convolutional autoencoder gets on one-class anomaly detection compared to classical shallow methods (PCA, LOF, CBLOF, KNN), and how much latent-space size and reconstruction loss choice matter along the way. Everything is trained in the one-class setting — only normal samples during training, anomalies show up only at test time as reconstruction failures.

Results (AUC-ROC)

Dataset	Autoencoder	Best Baseline	Baseline Method
MNIST	0.999	0.377	PCA
Fashion-MNIST	0.866	0.560	PCA
CIFAR-10	0.829	0.740	LOF
SVHN	0.631	0.596	LOF
MVTec-AD	0.483	0.653	PCA

MVTec-AD is the interesting exception — PCA beats the autoencoder there, most likely because resizing the industrial defect images down to 32×32 throws away exactly the fine texture the model needs to catch small anomalies. We left that result in rather than sweep it under the rug; it's a real finding about where this architecture stops working.

The live demo currently serves the MNIST, Fashion-MNIST, and CIFAR-10 checkpoints. SVHN is trained on paper but not uploaded yet — the training script hits a protobuf/tensorflow_datasets version conflict that needs a clean environment to sort out (tracked in #7).

Repository contents

• Image Data.ipynb — the experiments notebook: training, evaluation, and baseline comparisons
• demo/app.py — the Gradio app behind the HuggingFace Space
• demo/README_HuggingFace.md — Space config/README (deployed as the Space's README.md)
• demo/train_and_upload.py, demo/train_remaining_models.py — retrain a model and push its weights + threshold to the HF Hub
• demo/requirements.txt — pinned dependencies for the demo app
• .github/workflows/keep-space-awake.yml — pings the Space every 20 minutes so it doesn't fall asleep on HF's free tier

Publication

• Venue: IEEE IJCNN 2024
• Title: "Autoencoder Optimization for Anomaly Detection: A Comparative Study with Shallow Algorithms"
• DOI: 10.1109/IJCNN60899.2024.10650057
• Authors: Vikas Kumar · Vishesh Srivastava · Sadia Mahjabin · Arindam Pal · Simon Klüttermann · Emmanuel Müller

How to Use

Run the notebook locally

1. Clone the repository:

   git clone https://github.com/iamvisheshsrivastava/Autoencoder-Optimization-Anomaly-Detection.git

2. Create a Python environment with the required libraries:

   pip install numpy pandas tensorflow scikit-learn matplotlib jupyter

3. Open Image Data.ipynb in Jupyter Notebook or JupyterLab
4. Update any dataset paths or environment-specific settings before running all cells

Run the demo app locally

cd demo
pip install -r requirements.txt
python app.py

Citation

@inproceedings{Kumar2024Autoencoder,
  title={Autoencoder Optimization for Anomaly Detection: A Comparative Study with Shallow Algorithms},
  author={Kumar, Vikas and Srivastava, Vishesh and Mahjabin, Sadia and Pal, Arindam and Klüttermann, Simon and Müller, Emmanuel},
  booktitle={Proceedings of the International Joint Conference on Neural Networks (IJCNN)},
  year={2024}
}

License

This project is licensed under the MIT License. See LICENSE.