🌟Deep Learning Approaches for Deepfake Detection🌟

A powerful and comprehensive deepfake detection system implementing multiple deep learning architectures, all packed into a user-friendly web interface! 🚀

🌐 Live Web Application: Deepfake Detection Web App

---

🔍 Project Overview

This project integrates cutting-edge deep learning architectures to detect deepfakes:

• 🤖 Xception
• ⚡ EfficientNet
• 🌀 Swin Transformer
• 🔗 Cross Attention
• 🧠 CNN-Transformer

Each model is independently trained but seamlessly combined in the web app for ensemble predictions.

---

Example of the interactive web app interface

---

Results from multiple models displayed together

---

Image mode with face detection visualization

---

Video mode displaying prediction results

---

Cropped face samples from video

---

📜 License

This project is licensed under the Apache License 2.0. You are free to use, modify, and distribute the code, provided that the following conditions are met:

1. Attribution: Proper credit must be given to the original authors, along with a copy of the license.
2. State of Derivative Works: Any modifications or derivative works must be clearly identified as such.
3. No Liability: The software is provided "as is," without any warranties or guarantees. The authors are not liable for any damages arising from its use.

For full terms and conditions, refer to the official LICENSE.md file.

---

📊 Dataset

The project utilizes a balanced dataset of 20,000 face-cropped images:

• 🟢 10,000 Real Images
• 🔴 10,000 Fake Images
• 📚 Sourced from DFDC, FF++, and CelebDF-v2

📥 Dataset Link: 3body-filtered-v2-10k

---

🛠️ Training the Models

📋 Prerequisites

• A Kaggle account with GPU access
• Add the dataset to your Kaggle account

🔧 Training Steps

1. Use the provided train_*.py files in Kaggle notebooks
2. Each model has a unique training configuration
3. General training parameters:

   NUM_EPOCHS = 30  # Default
   BATCH_SIZE = 32  
   IMAGE_SIZE = 224  # Varies by model

4. Key configurations:
   • Optimizer: AdamW with weight decay
   • Scheduler: ReduceLROnPlateau
   • Experiment tracking: MLflow integration

🔄 Model Conversion

Models are converted for CPU compatibility:

• Use the notebook: DDP-V4-Converter
• Download pre-converted models: DDP-V4-Models

---

🌐 Web Application Setup

⚙️ Environment Setup

1. Create a virtual environment:

   python -m venv venv  
   source venv/bin/activate  # Linux/Mac  
   venv\Scripts\activate     # Windows  

2. Install dependencies:

   pip install -r requirements.txt  

📂 Application Structure

deepfake-detection-project-v4/  
├── app.py                   # Main Streamlit app  
├── video_processor.py       # Video utilities  
├── feature_visualization.py # Visualize CNN features  
├── data_handler.py          # Data processing  
├── train_*.py               # Training scripts  
└── converted_models/        # Directory for pre-trained models  

▶️ Run the Application

1. Add converted models to the converted_models/ folder
2. Start the app:

   streamlit run app.py  

---

🖥️ Using the Web Application

🖼️ Image Analysis

• Upload an image containing a face
• View:
  • 📸 Face detection visualization
  • 📊 Predictions from each model
  • 🧠 Feature maps (CNN models only)

🎥 Video Analysis

• Upload a video file
• Choose the number of frames to analyze (10–300)
• View:
  • 🔍 Frame-by-frame predictions
  • 😃 Detected faces
  • 📈 Confidence scores and stats

---

💡 Features

• 🔗 Multi-model ensemble predictions
• 🕵️‍♂️ Real-time face detection
• 🖼️ Feature map visualization
• 📊 Confidence tracking
• 🌑 Dark theme UI

---

🏗️ Model Architecture Details

Xception 🤖

• Specialized in texture analysis
• Input size: 299x299

EfficientNet ⚡

• B3 variant: Efficiency-accuracy balance
• Input size: 300x300

Swin Transformer 🌀

• Hierarchical feature learning
• Input size: 224x224

Cross Attention 🔗

• Focuses on enhanced feature interaction
• Input size: 224x224

CNN-Transformer 🧠

• Hybrid model: Combines local & global feature learning
• Input size: 224x224

---

📋 Requirements

• Python 3.8+
• CUDA-compatible GPU (for training)
• CPU with 8GB+ RAM (for inference)

---

📝 Notes

• ⚠️ Models must be converted before use in the web app
• 🖼️ Feature visualization is limited to CNN-based models
• ⏳ Video processing speed depends on frame count