🔍 Deepfake Image Detection with SqueezeNet

This project uses PyTorch and SqueezeNet to train a deepfake image classifier. It distinguishes between real and fake (deepfake) face images using transfer learning and mixed-precision training for efficiency.

---

📁 Project Structure

.
├── data/
│   ├── train/
│   │   ├── real/
│   │   └── fake/
│   └── val/
│       ├── real/
│       └── fake/
├── squeezenet_model.pth        # Saved trained model
├── main.py                     # Main training & inference script
└── README.md

---

🚀 Features

• Uses SqueezeNet for lightweight performance
• Transfer learning with pre-trained weights
• Mixed Precision Training via torch.amp for faster computation
• Modular structure for easy use in training and inference
• Automatically saves the best-performing model based on validation accuracy

---

📦 Requirements

• Python 3.8+
• PyTorch
• torchvision
• tqdm
• PIL

Install dependencies:

pip install torch torchvision tqdm pillow

---

🧠 Dataset Format

Organize the dataset in this format:

data/
├── train/
│   ├── real/
│   └── fake/
└── val/
    ├── real/
    └── fake/

Each folder should contain .jpg or .png image files.

---

🏋️‍♂️ Training

To train the model:

python main.py

• The best model is saved as squeezenet_model.pth
• Default settings:
  • Epochs: 25
  • Batch size: 32
  • Optimizer: Adam
  • Learning rate: 1e-5

---

🔍 Inference

Update the image path in main.py:

image_path = 'data/val/real/real_1.jpg'

After training, the model will predict the class of this image and print it:

Prediction: real

---

⚙️ Customization

You can modify:

• The model architecture (e.g., ResNet, MobileNet)
• Image size or augmentations
• Data directories and batch sizes
• Learning rate and optimizer in main.py

---

💡 Notes

• Uses SqueezeNet 1.0, a compact CNN architecture (~5MB)
• Fine-tuned last layer (Conv2d) for binary classification
• Uses AMP (Automatic Mixed Precision) for performance on GPU

---

📜 License

This project is open-source and available under the MIT License.