🛡️ Adversarial Face Recognition Attacks

This project demonstrates how to spoof a face recognition model using adversarial perturbations. Using facenet-pytorch, we simulate a real-world biometric attack scenario where one person's face is subtly altered to match the identity of another.

🌐 Overview

• Model: FaceNet (InceptionResnetV1, pretrained on VGGFace2)
• Framework: PyTorch
• Attack Method: Gradient-based cosine similarity optimization
• Defenses: JPEG Compression, Gaussian Blur

---

⚔️ Objective

To craft an adversarial example of Person B that is misclassified by the model as Person A, despite the two being visually distinct.

🧰 Project Structure

adversarial-face-spoofing/
├── face_recognition_model.py   # Embedding + similarity logic
├── adversarial_attack.py       # Spoof generation and evaluation
├── defense.py                  # Input transformations (JPEG, blur)
├── images/                     # Input face images
│   ├── person_a.jpg
│   ├── person_b.jpg

---

📊 Attack Results

Step	Cosine Similarity	Comment
0	0.18	Low initial similarity
100	1.00	Perfect spoof

Final similarity scores:

Adversarial image:   1.0000
JPEG compressed:     0.4812
Gaussian blurred:    0.5377

✅ Defenses were successful: similarity dropped below 0.6 threshold.

---

🔒 How the Attack Works

1. Extract embeddings of Person A and Person B
2. Iteratively adjust Person B's pixels to match Person A's embedding
3. Use cosine similarity loss and optimize using Adam
4. Clamp pixel values to stay within valid image bounds

---

🛡️ Defenses Implemented

• JPEG Compression: Removes minor pixel-level noise by re-encoding the image
• Gaussian Blur: Smooths input to degrade high-frequency adversarial perturbations

---

⚡ Setup Instructions

☝ Important

Use Python 3.10. Newer versions (like 3.12+) may break compatibility with facenet-pytorch or torch.

🔧 Create Environment and Install Requirements

python3.10 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

---

🔧 Run the Attack

python adversarial_attack.py

---

🌐 Author

Built by Gregory Apostle as part of an AI Security portfolio project.