AI vs. Human Face Detection

📌 Project Overview

This repository contains two deep learning models with identical architectures but different training strategies for distinguishing AI-generated faces from real human photographs. The models achieve state-of-the-art performance in detecting synthetic facial imagery with exceptional accuracy and generalization capabilities.

🎯 Key Features

• Dual Model Approach: Two PyTorch models with same architecture, different training strategies
• High Accuracy: 99.2% test accuracy without overfitting
• Real-world Validation: Successfully tested on real smartphone camera images and AI-generated faces
• Perfect Detection: 100% accuracy on validation set with real-world samples
• Comprehensive Dataset: Mixed sources of real and synthetic facial images

📊 Performance Metrics

Metric	Model 1	Model 2	Combined Evaluation
Test Accuracy	99.2%	99.1%	-
Real Images Detection	-	-	100%
AI-generated Detection	-	-	100%
Validation Accuracy	99.0%	98.8%	-

🏗️ Architecture

Both models share the same neural network architecture:

• Backbone: Custom CNN architecture (details in model.py)
• Input Size: 224×224 RGB images
• Output: Binary classification (Real vs AI-generated)
• Loss Function: Cross-entropy with regularization
• Optimizer: Adam with cosine annealing scheduler

🚀 Training Strategies

Model 1: Standard Training

• Standard data augmentation
• Balanced class weighting
• Early stopping implementation

Model 2: Enhanced Training

• Advanced augmentation pipeline
• Curriculum learning approach
• Focal loss implementation
• Mixed precision training

📂 Repository Structure

├── models/
│   ├── model_1.pt
│   ├── model_2.pt
├── dataset/
│   ├── dataset_real_or_fake_face.rar
├── notebook/
│   ├── notebook.ipynb
├── README.md
└── LICENSE

📊 Dataset Details

The dataset consists of:

• 1290 Face Images
• 590 Real Face Images
• 700 AI Generated Face Images

📝 Citation

If you use this work in your research, please cite:

@software{AIGenerated_Or_Real_Image_Detection,
  title = {AI vs Human Face Detection: Dual Model Approach},
  author = {Javad Rahimi},
  year = {2025},
  url = {https://github.com/AloneMaster7/AIGenerated_Or_Real_Image_Detection}
}

📄 License

This project is licensed under the MIT License.

📧 Contact

For questions or collaborations, please open an issue or contact:

• Javad Rahimi - jvd.r.403@gmail.com
• Telegram ID - @AlServiceHub

---

⭐ Star this repository if you find it useful!