🧠 MambaVision-Based Biometric Verification & Vulnerability Analysis

📌 Overview

This project implements a biometric identity verification system using a MambaVision backbone trained on forehead images. The system evaluates identity similarity using cosine similarity of learned embeddings and analyzes robustness against adversarial (E-dataset) inputs.

---

🚀 Key Features

• MambaVision-based feature learning (fallback mode)
• Embedding-based identity representation (640-D vectors)
• Cosine similarity for verification
• Genuine vs Imposter comparison
• Attack (E-dataset) vulnerability analysis
• End-to-end pipeline: training -> feature extraction -> evaluation

---

📂 Dataset Structure

Each subject contains 20 images:

Subject_i/
  ├── 1.jpg
  ├── ...
  ├── 10.jpg   -> Gallery (reference)
  ├── 11.jpg
  ├── ...
  ├── 20.jpg   -> Probe (query)

• Total subjects: 290
• Total images: 5800

---

⚙️ Pipeline

1) Model Training

• Backbone: MambaVision (fallback mode)
• Input: 224x224 images
• Epochs: 25
• Optimizer: Adam
• Scheduler: Learning rate decay applied

Output:

• Trained model: best_mamba_model.pth

---

2) Feature Extraction

Each image is converted into a:

640-dimensional embedding vector

Command:

python extract_features.py

Output:

features.pt

---

3) Similarity Computation

Cosine similarity is computed between:

• Genuine pairs (same subject)
• Imposter pairs (different subjects)

Pairing strategy:

Gallery (1-10) vs Probe (11-20)

Command:

python similarity.py

---

4) Attack Evaluation (E-dataset)

Evaluate robustness using modified images:

Real Gallery vs E-dataset Probe

Command:

python extract_features_e.py
python attack_similarity.py

---

📊 Results

Type	Avg Similarity
Genuine	~0.681
Imposter	~0.0108
Attack	~0.340

---

📈 Plots

Loss Curve

Validation Accuracy Curve

Similarity Comparison

---

📈 Interpretation

• High genuine similarity -> strong identity representation
• Low imposter similarity -> good class separation
• Moderate attack similarity -> partial vulnerability

The model remains robust but is not fully resistant to adversarial inputs.

---

🧠 Key Insights

• Embedding-based verification is more flexible than classification
• Cosine similarity effectively separates identities
• Attack data reduces similarity but does not collapse identity recognition
• Verification requires full subject data, not split datasets

---

🔮 Future Work

• ROC / FAR / FRR analysis
• Threshold-based decision system
• Grad-CAM visualization
• Robust training against adversarial samples
• Deployable verification system

---

💡 How to Run

# Train model
python train.py

# Extract features (real dataset)
python extract_features.py

# Compute similarity
python similarity.py

# Extract features (E-dataset)
python extract_features_e.py

# Attack evaluation
python attack_similarity.py

---

📬 Contact / Notes

This project is part of a Deep Learning study on biometric verification systems. Feedback and suggestions are welcome.