DASVDD

Implementation of DASVDD (Deep Autoencoding Support Vector Data Descriptor) for anomaly detection and one-class classification.

The project combines dataset-specific autoencoders with an SVDD-inspired objective to learn compact representations of normal data and score anomalies through reconstruction quality and distance to a learned center.

• Supports MNIST, FMNIST, CIFAR, Speech, and PIMA
• Includes dataset-specific autoencoder architectures
• Tunes the SVDD weighting term before training
• Uses deterministic preprocessing for reproducible runs
• Provides a simple command-line interface for running experiments

.
├── main.py               # CLI entry point
├── src/
│   ├── core/             # Training, evaluation, and gamma tuning
│   ├── data/             # Dataset loaders and preprocessing helpers
│   └── models/           # Dataset-specific autoencoders
├── data/                 # Local tabular datasets
└── requirements.txt      # Python dependencies

git clone https://github.com/Armanfard-Lab/DASVDD.git
cd DASVDD

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Run DASVDD on any supported dataset:

python3 main.py --dataset MNIST --target_class 0 --epochs 30 --batch_size 128

Example runs:

python3 main.py --dataset FMNIST --target_class 3 --epochs 20
python3 main.py --dataset CIFAR --target_class 1 --epochs 10
python3 main.py --dataset PIMA --epochs 30
python3 main.py --dataset Speech --epochs 30

For Speech and PIMA, --target_class is ignored because the datasets already include anomaly labels.

• MNIST and FMNIST are downloaded automatically through torchvision.
• CIFAR is downloaded automatically and normalized with global contrast normalization.
• Speech and PIMA are loaded from the local data/ directory.
• Tabular datasets are shuffled with a fixed seed before splitting so evaluation is reproducible without depending on CSV row order.

• The refactored modules expose clearer APIs under src.core, src.data, and src.models.
• Backward-compatible aliases such as DASVDD_trainer, DASVDD_test, and AE_MNIST are still available for older notebooks or scripts.

The CLI prints progress for:

• gamma tuning
• DASVDD training
• final ROC-AUC evaluation

If you use this repository in academic work, please cite:

@ARTICLE{DASVDD,
  author={Hojjati, Hadi and Armanfard, Narges},
  journal={IEEE Transactions on Knowledge and Data Engineering},
  title={DASVDD: Deep Autoencoding Support Vector Data Descriptor for Anomaly Detection},
  year={2024},
  volume={36},
  number={8},
  pages={3739-3750},
  keywords={Anomaly detection;Training;Task analysis;Support vector machines;Image reconstruction;Data models;Benchmark testing;Anomaly detection;deep autoencoder;deep learning;support vector data descriptor},
  doi={10.1109/TKDE.2023.3328882}
}

Paper link: IEEE Xplore