🛡️ Adversarial ML Toolkit

This repository provides implementations of various adversarial attacks and their corresponding defense mechanisms on machine learning models. It is structured for easy experimentation and modular evaluation.

---

📌 Implemented Attacks & Defenses

Attack Type	Description	Defense Mechanism
FGSM Attack	Fast Gradient Sign Method	Adversarial Training
FALFA Attack	Feature Alignment-based Label Flipping	kNN-based Defense
Backdoor Attack	Universal Trigger-based Backdoor Attack	Spectral Signature & Pruning-based Defense

---

🧪 How to Use

🔍 Run Evaluation Notebook

To evaluate all attack and defense methods in one place:

# Open the notebook in JupyterLab, VS Code, or Google Colab
evaluation.ipynb

This notebook includes visualizations, performance metrics, and comparisons of model accuracy before and after applying defenses against attacks.

---

🧪 Run Individual Pipelines

Each attack-defense experiment is organized into separate Python scripts under the testing/ directory.

Example usage:

cd testing/

Available Pipelines:

# Run FGSM Attack with Adversarial Training Defense
python test_fgsm_attack_defense.py

# Run FALFA Attack with kNN-based Defense
python test_knn_defense.py

# Run Backdoor Attack with Spectral Signature + Pruning Defense
python test_pruning.py | python test_ss.py

---

📂 Project Structure

.
├── evaluation.ipynb                # Main evaluation notebook
├── testing/                        # Pipelines for attacks & defenses
│   ├── test_fgsm_attack_defense.py
│   ├── test_knn_defense.py
│   ├── test_pruning.py
│   └── ...
├── attacks/
├── defences/
├── save_path/
├── data.csv
├── models/                         # Neural network model definitions
├── data/                           # Sample dataset (or link to external source)
├── utils/                          # Utility functions for attacks/defenses
└── README.md                       # Project overview

---

📚 References

• FGSM – Fast Gradient Sign Method: 1Konny/FGSM
• Backdoor Attack & Defense – BATD: HamidRezaTajalli/BATD
• Spectral Signature Defense – Detecting Backdoor Attacks on Deep Neural Networks by Activation Clustering, Tran et al., 2018

Meomeo <3

---

⚠️ Disclaimer

This project is for educational and research purposes only. Any use of adversarial attacks for malicious purposes is strictly discouraged.

---