Lightweight DeepFake Detection System

A lightweight DeepFake detection system built using PyTorch, OpenCV, and Gradio. The project detects whether an image frame (or video frame) is REAL, FAKE, or UNCERTAIN based on model confidence.

---

🚀 Features

• ⚡ Lightweight CNN-based DeepFake detector
• 🖼️ Image-based inference (video frames supported)
• 📊 Confidence-aware predictions (REAL / FAKE / UNCERTAIN)
• 🌐 Interactive web UI using Gradio
• 🧪 Dummy dataset generation for quick testing

---

📸 Project Demo (UI Screenshots)

Left → Right: Upload Interface • Image Preview • Prediction Result with Confidence

---

🗂️ Project Structure

Lightweight-DeepFake-Detection-System/
│
├── models/
│   └── deepfake_cnn.pth              # Trained CNN model weights
│
├── src/
│   ├── data/
│   │   ├── create_dummy_data.py      # Generates sample training data
│   │   ├── dataset.py                # PyTorch Dataset class
│   │   ├── frame_extractor.py        # Extracts frames from videos
│   │   ├── preprocessing.py          # Image preprocessing utilities
│   │   └── run_frame_extraction.py   # Script to extract frames
│   │
│   ├── train/
│   │   ├── train.py                  # Model training logic
│   │   └── evaluate.py               # Model evaluation script
│   │
│   ├── inference/
│   │   ├── app.py                    # Gradio web application
│   │   ├── predict_image.py          # Image-based inference
│   │   └── predict_video.py          # Video-based inference
│   │
│   └── __init__.py
│
├── quick_train.py                    # Fast training script (demo use)
├── requirements.txt                  # Project dependencies
├── README.md                         # Project documentation
└── .venv/                            # Virtual environment (ignored)

---

🧠 Model Architecture

A lightweight Convolutional Neural Network (CNN) consisting of:

• 2 × Convolutional layers
• ReLU activation functions
• MaxPooling layers
• Fully connected classifier (2 classes: REAL / FAKE)

🎯 Design Goals

• Fast inference
• Low computational overhead
• Easy deployment on low-resource systems

---

🧪 Prediction Logic

The model outputs probabilities for both classes:

• REAL probability
• FAKE probability

Decision Rules

• REAL ✅ → confidence > threshold
• FAKE ❌ → confidence > threshold
• UNCERTAIN 🤔 → confidence below threshold

⚠️ The UNCERTAIN class is intentional to reduce false positives when the model is not confident.

---

📌 Use Cases

• DeepFake detection in facial images
• Educational demos for AI & Computer Vision
• Internship / hiring challenge submissions
• Research prototypes for media authenticity

---

🔮 Future Improvements

• Full video-level DeepFake detection
• Advanced architectures (EfficientNet / Vision Transformers)
• Explainable AI (Grad-CAM visualizations)
• Cloud deployment

---

👤 Author

Samta Sakha
Data Science & AI Enthusiast