🔍 Enhanced Deepfake Detection System with Ensemble Models

📁 Project Structure

.
├── prepare\_cifar\_binary.py          # Converts CIFAR-10 to binary real/fake dataset
├── multi\_model\_inference.py        # Training, evaluation, and inference pipeline
├── cifar\_binary/                   # Generated dataset directory (created automatically)
├── enhanced\_deepfake\_ensemble.pth  # Saved ensemble model (after training)
├── evaluation\_results/             # Metrics, plots, and reports
└── sample\_image.jpg                # Optional test image for demo inference

---

✅ Features

• 🔄 Ensemble of models: EfficientNet, ResNet, Vision Transformer
• 🧠 Face extraction option: Built-in OpenCV face detection support
• ⚖️ Handles class imbalance: Weighted sampling during training
• 📊 Detailed evaluation: AUC, ROC, Precision-Recall, Confusion Matrix
• 🧪 Single image inference: Predicts with ensemble and confidence score
• 🗂️ Synthetic dataset preparation: From CIFAR-10 (real/fake classes)

---

⚙️ Setup

🔧 Install Dependencies

pip install torch torchvision timm transformers scikit-learn matplotlib seaborn tqdm opencv-python

---

📦 Step 1: Prepare Dataset from CIFAR-10

1. Download CIFAR-10 Python version.
2. Update the cifar_dir variable in prepare_cifar_binary.py with the path to your dataset.
3. Run:

python prepare_cifar_binary.py

This creates the dataset in the following structure:

cifar_binary/
├── train/
│   ├── real/
│   └── fake/
├── val/
├── test/

Label Mapping:

• Real (Label 0): cat, deer, dog, horse
• Fake (Label 1): airplane, automobile, ship, truck

---

🚀 Step 2: Train the Deepfake Detection Ensemble

python multi_model_inference.py

This will:

• Train EfficientNet, ResNet, and ViT models
• Save best weights per model (e.g., best_resnet50.pth)
• Save the final ensemble as enhanced_deepfake_ensemble.pth
• Generate evaluation plots and metrics in evaluation_results/

---

🔍 Step 3: Inference on a Sample Image (Optional)

Place a file named sample_image.jpg in the project root and re-run the script:

python multi_model_inference.py

You will see per-model and ensemble predictions with confidence scores.

---

📈 Evaluation Output

• evaluation_results/evaluation_report.json: Accuracy, AUC, classification report
• evaluation_results/evaluation_plots.png: Confusion matrix, ROC, PR curves

---

💡 Notes

• Face extraction is optional and handled via OpenCV's Haar cascade.
• Early stopping and learning rate scheduling are included.
• The ensemble uses a weighted average strategy by default.

---

📜 License

For educational and research purposes only.