Facial_Deepfake-video_detection

Deepfake detection using CNN + LSTM models, served through Flask.

Repository: https://github.com/Shivam-26102003/Deepfake_facial-video_detection

---

🔖 Table of Contents

🔖 Table of Contents

• About
• Features
• Technologies
• Installation
• Usage
• Project Structure
• Models
• Web Application
• Results
• Contributing
• License

---

🧐 About

This repository implements a Facial Deepfake Detection pipeline combining:

• ResNeXt-50 CNN to extract spatial features from video frames.
• LSTM RNN to model temporal inconsistencies across frame sequences.
• A balanced multi-source dataset (FaceForensics++, DFDC, Celeb-DF).
• A Flask-based web interface for real-time video uploads and predictions.

---

📁 Dataset

This project uses a combination of publicly available deepfake datasets:

• 📦 FaceForensics++ – A popular dataset for facial manipulation detection.
• 📦 Deepfake Detection Challenge Dataset (DFDC) – Released by Facebook via Kaggle.
• 📦 Celeb-DF (v2) – High-quality deepfake dataset for research.

---

✨ Features

• Spatial Feature Extraction: ResNeXt-50 pretrained on ImageNet for robust embeddings.
• Temporal Analysis: LSTM captures frame-to-frame artifacts indicative of deepfakes.
• Web UI: Upload any video via browser and receive confidence scores instantly.
• Modular Design: Separate scripts for preprocessing, training, inference, and serving.

---

🛠 Technologies

• Language: Python 3.8+
• Deep Learning: PyTorch >=1.7
• Web Framework: Flask
• Computer Vision: OpenCV, face_recognition
• Data Handling: NumPy, Pandas
• Visualization: Matplotlib

---

⚙ Installation

1. Clone the repo bash git clone https://github.com/Shivam-26102003/Deepfake_facial-video_detection.git cd Deepfake_facial-video_detection

2. Set up a virtual environment bash python3 -m venv venv source venv/bin/activate # macOS/Linux venv\Scripts\activate # Windows

3. Install dependencies bash pip install -r requirements.txt

4. Prepare models Download pretrained ResNeXt-50 via PyTorch Hub. Place trained LSTM checkpoint in models/ directory.

▶ Usage

1. Preprocess videos bash python scripts/preprocess.py --input_dir data/raw --output_dir data/processed
2. Train the model bash python train.py --config configs/train.yaml
3. Inference on a video bash python predict.py --video path/to/video.mp4
4. Start the web server bash python server.py

📂 Project Structure

bash
  Deepfake_facial-video_detection/
  ├── data/                   # raw and processed video datasets
  ├── models/                 # model checkpoints
  ├── scripts/                # preprocessing & training scripts
  │   ├── preprocess.py
  │   └── train.py
  ├── static/                 # CSS, JS, images for Flask app
  ├── templates/              # HTML templates for Flask app
  ├── server.py               # Flask server entry-point
  ├── requirements.txt
  ├── LICENSE
  └── README.md

🔍 Models

1. ResNeXt-50_32x4d: Extracts 2048-dimensional feature vectors per frame.
2. LSTM: Single-layer, hidden size 2048, dropout 0.4, followed by a classifier.
3. Classifier: FC → LeakyReLU → Softmax for binary real/fake prediction.

🌐 Web Application

1. Flask backend (server.py) accepts video uploads and returns predictions.
2. Templates: Simple HTML form for file upload and result display.
3. Static: CSS for styling and JS for frontend interactions.

📈 Results

Performance on Mixed Dataset (6000 videos):

Metric Value

1. Accuracy 87.8%
2. Precision 89.3%
3. Recall 86.5%
4. F1-score 87.9%

🤝 Contributing

We welcome contributions to this project!
Special thanks to my teammate:

• Abhinav Patel
  LinkedIn

• Shivam-26102003

• LinkedIn

Feel free to open Issues or Pull Requests.

License

Distributed under the MIT License. See LICENSE for more information.