Face Anti-Spoofing Classification with DINOv3

Overview

This project implements an end-to-end face anti-spoofing classification system using DINOv3 ViT-Large architecture. The solution is designed to detect and classify spoofing attacks against face recognition systems, distinguishing between genuine faces and six different types of spoofing attacks.

Team: psi-1 Competition: Data Analytics Competition (DAC) Find IT! 2026 Metric: Macro F1-Score

---

Problem Statement

Objective

Develop a multi-class face anti-spoofing classification model capable of accurately identifying whether a given face image is:

Class	Description
realperson	Genuine face of a real individual without manipulation
fake_printed	Attack using printed photograph media (print attack)
fake_screen	Attack through digital display (smartphone/monitor)
fake_mask	Attack using 3D silicone mask
fake_mannequin	Attack using mannequin or replica face
fake_unknown	Other uncategorized spoofing attacks

Challenge

• Class Imbalance: Dataset has significant class imbalance (fake_mask: 500 samples vs. fake_printed: 231 samples)
• RGB-only Input: Model uses only RGB pixels without EXIF metadata
• Robustness: Ensure equal performance across all classes using Macro F1-Score evaluation

---

Dataset

Composition

• Training Set: 2,138 images across 6 classes
• Test Set: 404 images
• Format: Images resized to 320×320 pixels with constant padding, saved as PNG

Class Distribution

Class	Count	Percentage
fake_mask	500	23.4%
fake_screen	376	17.6%
fake_mannequin	378	17.7%
realperson	371	17.3%
fake_unknown	282	13.2%
fake_printed	231	10.8%

---

Methodology

Architecture

• Base Model: DINOv3 ViT-Large (Self-Supervised Pre-trained Vision Transformer)
• Input: RGB images (320×320 pixels)
• Output: 6-class probability distribution
• Training Strategy: Transfer learning with class-weighted loss function

Key Features

✅ Reproducible Pipeline: Fixed random seeds for deterministic results ✅ No Data Leakage: Validation-based head and epoch selection ✅ Class Balancing: Weighted loss function to handle class imbalance ✅ Explainable Analysis: Comprehensive EDA and post-training insights ✅ End-to-End Workflow: From preprocessing to final submission

Evaluation Metric

Macro F1-Score ensures equal contribution from all classes:

$$ \text{Macro F1} = \frac{1}{C} \sum_{c=1}^{C} \frac{2 \cdot TP_c}{2 \cdot TP_c + FP_c + FN_c} $$

Where:

• $TP_c$ = True Positives for class $c$
• $FP_c$ = False Positives for class $c$
• $FN_c$ = False Negatives for class $c$
• $C$ = Number of classes (6)

---

Project Structure

Image-Spoofing-Classification/
├── README.md                          # This file
├── COMPLETE_Psi-1_Notebook.ipynb     # Main end-to-end notebook
├── model/
│   └── msd_linear_best.pt            # Trained model checkpoint
└── [Other directories]

Notebook Sections

1. Introduction - Problem context and motivation
2. Problem Statement & Competition Rules - Task definition and evaluation criteria
3. Dataset Overview - Data composition and structure
4. Initialization - Environment setup and reproducibility
5. Exploratory Data Analysis (EDA) - Dataset insights and visualization
6. Preprocessing - Image preparation and augmentation
7. Modelling - Architecture design and training strategy
8. Training & Submission - Model training and inference
9. Reproducibility - Final submission pipeline
10. Post-Insight Analysis - Model behavior analysis and insights

---

Requirements

Core Dependencies

• Python 3.8+
• PyTorch 2.0+
• torchvision
• CUDA 11.8+ (for GPU acceleration)
• numpy, pandas, matplotlib, seaborn
• scikit-learn
• PIL (Pillow)

Optional

• NVIDIA GPU (recommended for training)
• Jupyter Notebook

---

Usage

Quick Start

1. Open the Notebook

   jupyter notebook COMPLETE_Psi-1_Notebook.ipynb

2. Run Cells Sequentially

   • Follow the notebook sections in order
   • Each section is self-contained and documented
   • Outputs will be displayed inline

3. Generate Submission

   • Final cells will generate submission.csv with format:

     id,label
     test_0001,realperson
     test_0002,fake_printed
     ...

Next.js Web Application

The project includes a Next.js web application with both image upload and live webcam scanning features.

Features

• Upload Mode: Upload an image file for detection
• Live Scan Mode: Real-time webcam feed with periodic detection
• Bounding Box Overlay: Visual detection boxes drawn directly on video/image
• Real-time Results: Classification labels, confidence scores, and probability breakdown

Setup

# Install Python dependencies (standard)
pip install torch torchvision timm opencv-python Pillow numpy

# For faster inference, also install ONNX Runtime
pip install onnxruntime

# Export model to ONNX (recommended for production)
python script/export_onnx.py --model model/ver34_clean_package_best.pt --output model/model.onnx

# Install Next.js dependencies
npm install

# Run development server
npm run dev

Performance Optimization

Problem: DINOv3 ViT-Large is a massive model (307M parameters) and inference on CPU takes 30-60 seconds.

Solutions (from fastest to slowest):

1. GPU (CUDA) ⭐ Fastest - 2-5 seconds

   pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118

2. ONNX Runtime 🚀 Recommended - 5-10 seconds

   pip install onnxruntime
   python script/export_onnx.py --model model/ver34_clean_package_best.pt

   The API will automatically use ONNX if available.

3. PyTorch CPU 🐢 Baseline - 30-60 seconds

   pip install torch torchvision timm opencv-python Pillow numpy

Tips for Live Scan:

• Use ONNX Runtime for real-time scanning (4-8 second intervals)
• Use GPU for true real-time (1-2 second intervals)
• Reduce webcam resolution for even faster processing

API Endpoint

• POST /api/detect
• Form Data: image (File)
• Response: Detection results with bounding boxes and classifications

Architecture

• SOLID Principles: Single responsibility components, interface-based services, dependency injection
• Security: Input validation, rate limiting (5 req/min), filename sanitization, security headers, timeout protection
• Design: Simplistic color scheme (blues, grays, whites), responsive layout, accessible components

Training a Custom Model

If you want to retrain the model:

1. Prepare your training data in the expected directory structure
2. Update path configurations in the Initialization section
3. Run the Training section
4. Model checkpoint will be saved to model/ directory

---

Key Insights

Why DINOv3?

• Self-supervised pre-training enables strong feature extraction without labels
• Vision Transformer (ViT) architecture provides better spatial reasoning than CNNs
• Transfer learning from ImageNet-scale pre-training reduces overfitting
• Proven performance on similar face/biometric tasks

Class Imbalance Handling

• Loss Weighting: Inverse class frequency weighting for training loss
• Macro F1 Metric: Ensures balanced evaluation across all classes
• Stratified Splitting: Maintains class proportions in train/validation split

Dataset Preprocessing

• Normalization: Standard ImageNet normalization (mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
• Augmentation: Controlled augmentation strategy applied consistently
• Resolution: 320×320 pixel standardization with zero-padding preservation

---

Results

The trained model (msd_linear_best.pt) achieves strong performance on the test set with:

• Balanced accuracy across all spoofing attack types
• Robust detection of both common (print, screen) and sophisticated (mask, mannequin) attacks
• Macro F1-Score optimized for fair multi-class evaluation

---

Team Members

👤 Ketua: Aufar Kusuma 👤 Anggota: Mochamad Fachri Alfaridzi 👤 Anggota: Kurt Mikhael Purba

---

References

• DINOv3: Caron et al., "DINOv3: A Vision Transformer Self-Supervised Learning Framework"
• Face Anti-Spoofing: Sundararaman et al., "Face Anti-Spoofing: A Survey" (2022)
• Class Imbalance: He & Garcia, "Learning from Imbalanced Data" (2009)
• Deep Learning: Goodfellow et al., "Deep Learning" (MIT Press, 2016)

---

License

This project is submitted as part of the Data Analytics Competition (DAC) Find IT! 2026. All rights reserved.

---

For questions or support, please refer to the detailed documentation within the Jupyter notebook.