🕵️ Deepfake Detector

A deep learning system for detecting AI-generated / manipulated media using EfficientNet-B0, trained with PyTorch Lightning and served through a Flask web app with a modern drag-and-drop UI.

---

✨ Features

• Binary classification — Real vs. Deepfake with confidence score
• Image & video support — JPG, PNG, MP4, MOV
• EfficientNet-B0 backbone — lightweight yet accurate
• Custom dataset loader — supports mixed data sources with optional label overrides
• Config-driven training — all hyperparameters in config.yaml
• Early stopping & LR scheduling — smart training callbacks
• Flask web interface — sleek dark UI with live confidence bar and media preview

---

📁 Project Structure

deepfake-detector/
│
├── datasets/
│   └── hybrid_loader.py       # Custom PyTorch Dataset with multi-source support
│
├── lightning_modules/
│   └── detector.py            # PyTorch Lightning module (train/val/optimizer)
│
├── models/
│   └── best_model-v3.pt       # Trained model weights (not included in repo)
│
├── web-app1.py                # Flask inference server + embedded UI
├── main_trainer.py            # Training entry point
├── config.yaml                # Hyperparameters and dataset paths
└── requirements.txt           # Python dependencies

---

⚙️ Setup

1. Clone the repository

git clone https://github.com/moheezkiani01/deepfake-detector.git
cd deepfake-detector

2. Install dependencies

pip install -r requirements.txt

CUDA users: Make sure your torch install matches your CUDA version.
Visit pytorch.org for the right install command.

3. Prepare your dataset

Your dataset directories should follow this structure:

C:/Datasets/your_dataset/
├── train/
│   ├── real/
│   │   ├── image1.jpg
│   │   └── ...
│   └── fake/
│       ├── image1.jpg
│       └── ...
└── validation/
    ├── real/
    └── fake/

Then update config.yaml with your actual paths:

train_paths:
  - C:/Datasets/your_dataset/train

val_paths:
  - C:/Datasets/your_dataset/validation

---

🚀 Training

python main_trainer.py

The best model checkpoint will be saved to models/best_model.ckpt.

Training configuration (config.yaml)

Parameter	Default	Description
lr	0.0001	Learning rate
batch_size	4	Batch size
num_epochs	1	Max training epochs
early_stopping_patience	3	Epochs to wait before stopping
scheduler_factor	0.5	LR reduction factor on plateau
scheduler_patience	2	Epochs to wait before reducing LR
monitor_metric	val_loss	Metric tracked by callbacks
save_top_k	1	Number of best checkpoints to keep

---

🌐 Running the Web App

Make sure you have a trained model saved at models/best_model-v3.pt.

python web-app1.py

Then open your browser at http://localhost:5000.

How it works

1. Drag & drop or browse for an image (JPG/PNG) or video (MP4/MOV)
2. Click Analyze File
3. The app returns a Real / Deepfake verdict with a confidence percentage and a preview of the analyzed frame

For video files, inference is performed on the first frame.

---

🧠 Model Architecture

Component	Details
Backbone	EfficientNet-B0 (ImageNet pretrained)
Classifier head	Dropout(0.4) → Linear(1280, 2)
Loss	Cross-Entropy
Optimizer	Adam
Input size	224 × 224
Output classes	2 (Real = 0, Fake = 1)

---

📦 Dependencies

torch
torchvision
pytorch-lightning
flask
opencv-python
PyYAML
Pillow
numpy
scikit-learn
tensorboard

---

📝 Notes

• The HybridDeepfakeDataset supports loading from multiple directories at once and allows per-source label overrides, making it easy to combine datasets (e.g., FaceForensics++, Celeb-DF, custom data).
• Training automatically uses GPU if available, falling back to CPU.
• Logs are available via TensorBoard: tensorboard --logdir lightning_logs/

---