🛡️ LunaShield: AI-Powered Deepfake Detection Framework

LunaShield is a comprehensive deepfake detection framework that leverages cutting-edge computer vision and deep learning to analyze, classify, and report on the authenticity of video content.

It is built for end-to-end forensic analysis of deepfake videos—right from frame extraction, model training, evaluation, and finally, PDF reporting with visualizations. Ideal for researchers, journalists, digital forensics teams, or developers who want a production-ready solution.

---

📌 Key Features

• 🎞️ Video Frame Extraction — Captures keyframes from real/fake videos
• 🧠 Deep Learning Engine — EfficientNet-B0 architecture pretrained on ImageNet
• 🧪 Advanced Data Splitting — Stratified train/val/test with augmentation
• ⚡ Mixed Precision Training — With torch.cuda.amp for optimized GPU usage
• 📉 Performance Monitoring — Real-time loss/accuracy + confusion matrix plots
• 📝 Automated Report Generation — Clean PDF with predictions + decision summary
• 🧰 Modular Codebase — Well-separated extraction, training, evaluation, reporting

---

📂 Directory Structure

├── dataset/                   # Original input videos (real/fake)
├── extracted frames/         # Extracted image frames
├── split_data/               # Organized train/val/test sets
├── report_charts/            # Plots: confusion matrix, bar graphs
├── Deepfake_Report.pdf       # Final analysis report
└── full_pipe.py              # Main execution pipeline

---

🧠 Model Architecture

• Backbone: EfficientNet-B0 (from torchvision)
• Custom Head: Dropout + Linear (for binary classification)
• Training Epochs: 10 (with early stopping)
• Loss Function: CrossEntropyLoss
• Optimizer: Adam (lr=3e-4)
• Scheduler: ReduceLROnPlateau (auto adjusts LR)
• Precision: Mixed (via AMP and GradScaler)
• Input Size: 224x224 RGB
• Classes: Real, Fake

---

🧪 Workflow Overview

1. 🔍 Video Preprocessing

extract_frames(VIDEO_DIR, EXTRACTED_FRAMES_DIR)

2. 🔀 Dataset Splitting

split_dataset(EXTRACTED_FRAMES_DIR, './split_data')

3. 📚 Training

train_model()

4. 📊 Evaluation

evaluate_model()

5. 📄 PDF Report Generation

generate_report("your_video.mp4")

---

🧱 Dependencies

Install the following Python packages:

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

Or use this requirements.txt:

torch
torchvision
opencv-python
matplotlib
seaborn
scikit-learn
reportlab
tqdm

---

📈 Example Report Summary

Metric	Value
Total Frames Analyzed	10
Real Frames Detected	6 (60%)
Fake Frames Detected	4 (40%)
Final Conclusion	REAL CONTENT

---

🛠️ Customization Guide

Task	How to do it
Change number of frames	Edit FRAMES_PER_VIDEO in full_pipe.py
Use your own videos	Replace files in dataset/real/ and dataset/fake/
Change model backbone	Swap EfficientNet with ResNet/VisionTransformer/etc.
Tune hyperparameters	Adjust BATCH_SIZE, NUM_EPOCHS, LEARNING_RATE, etc.
Improve augmentations	Edit data_transforms dictionary
Change report design	Modify the generate_report function using ReportLab

---

🔍 Technical Highlights

• Optimized for Speed
• Modular & Reusable
• Scalable
• Production-Ready

---

📌 Roadmap / TODOs

• Add Streamlit-based UI
• Real-time webcam detection
• LSTM/CNN hybrid for spatio-temporal detection
• Audio-video multimodal support
• FastAPI deployment

---

🤝 Contributing

Contributions are welcome!

---

📃 License

MIT License

---

⭐️ If you like this project...

Please star ⭐️ the repo and share!