GANFS: GAN-Based Feature Selection

A Python library for feature selection using Generative Adversarial Networks. GANFS trains a GAN on your data and uses perturbation-based sensitivity analysis on the discriminator to rank and select the most important features.

Installation

Install via PyPI (Recommended):

pip install ganfs

Install from GitHub (Latest Development Version):

pip install git+https://github.com/patelharsh15/GANFS-GAN-based-feature-selection.git

From source:

git clone https://github.com/patelharsh15/GANFS-GAN-based-feature-selection.git
cd GANFS-GAN-based-feature-selection
pip install -e .

Quick Start

from ganfs import GANFS
import pandas as pd

# Load your dataset
df = pd.read_csv("my_data.csv")
X = df.drop("label", axis=1)
y = df["label"]

# Initialize and train GANFS
selector = GANFS(epochs=200, batch_size=4096)
selector.fit(X, y)

# View feature ranking
ranking = selector.get_feature_ranking()
print(ranking)

# Select top 20 features
X_selected = selector.transform(X, k=20)

# Save/load trained models
selector.save("my_ganfs_model")
loaded = GANFS.load("my_ganfs_model")

API Reference

GANFS Class

Constructor Parameters

Methods

Usage with scikit-learn

from ganfs import GANFS
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import accuracy_score

# Feature selection
selector = GANFS(epochs=200)
selector.fit(X_train, y_train)
X_train_selected = selector.transform(X_train, k=20)
X_test_selected = selector.transform(X_test, k=20)

# Downstream classification
clf = RandomForestClassifier()
clf.fit(X_train_selected, y_train)
accuracy = accuracy_score(y_test, clf.predict(X_test_selected))
print(f"Accuracy with top-20 GANFS features: {accuracy:.4f}")

How It Works

1. GAN Training — A Generator-Discriminator pair is trained on the feature data. The Generator learns to produce realistic synthetic samples, while the Discriminator learns to distinguish real from fake.

2. Sensitivity Analysis — After training, each feature is perturbed (using dynamic perturbation magnitudes scaled to each feature's natural granularity) and the discriminator's response is measured. Features that cause the largest output changes are the most discriminative.

3. Feature Ranking — Features are ranked by their average sensitivity scores across multiple perturbation levels and directions.

4. Feature Selection — The top-K features can be selected for downstream tasks (classification, regression, etc.).

Project Structure

├── ganfs/                           # Python package source code
│   ├── __init__.py                  # Public API
│   ├── ganfs.py                     # Main GANFS class
│   ├── models.py                    # Generator & Discriminator networks
│   ├── sensitivity.py               # Sensitivity analysis functions
│   └── utils.py                     # GPU setup & preprocessing utilities
├── pyproject.toml                   # Package build configuration
└── research_archive/                # Original research files and experiments
    ├── GAN Algo Final.ipynb         # Original research notebook
    ├── benchmarking.ipynb           # Benchmarking vs traditional methods
    ├── training_checkpoints/        # Saved model checkpoints
    ├── feature_pair_interactions.csv
    └── feature_sensitivity_results.csv

Dataset Setup (for reproducing research results)

The original research uses the CIC-DDoS2019 dataset. The dataset files are too large (~12 GB) to host on GitHub.

Download Instructions

1. Visit the CIC-DDoS2019 dataset page
2. Request access and download the following CSV files:
   • DrDoS_DNS.csv, DrDoS_LDAP.csv, DrDoS_MSSQL.csv, DrDoS_NTP.csv
   • DrDoS_NetBIOS.csv, DrDoS_SNMP.csv, DrDoS_SSDP.csv, DrDoS_UDP.csv
3. Place all files in a CIC-DDoS2019/ folder at the repository root
4. Update the base_path in the notebook to "./CIC-DDoS2019/"

Requirements

• Python 3.8+
• TensorFlow 2.x (GPU support recommended)
• NumPy, Pandas, scikit-learn

Citation

If you use GANFS in your research, please cite our paper:

@misc{patel2025ganfs,
      title={GANFS: GAN-based Feature Selection for Machine Learning}, 
      author={Harsh Patel},
      year={2025},
      eprint={2504.18566},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2504.18566}, 
}

Acknowledgments

This project uses the CIC-DDoS2019 dataset provided by the Canadian Institute for Cybersecurity, University of New Brunswick.

If you use the dataset, please cite the original authors:

Iman Sharafaldin, Arash Habibi Lashkari, Saqib Hakak, and Ali A. Ghorbani,
"Developing Realistic Distributed Denial of Service (DDoS) Attack Dataset and Taxonomy",
IEEE 53rd International Carnahan Conference on Security Technology, Chennai, India, 2019.

License

MIT License — see LICENSE for details.