DeepFake Detection using VGG16 and SVM

🔍 Overview

• Model: Pretrained VGG16 from PyTorch, fine-tuned for feature extraction.
• Classifier: Support Vector Machine (SVM) for final classification based on extracted features.
• Web Interface: Streamlit app to upload an image and predict whether it's Real or Fake.
• Dataset: 140K Real and Fake Faces from Kaggle

📁 Project Structure

.
├── app.py
├── train.ipynb
├── model_15.pth
├── requirements.txt
└── README.md

📄 File Descriptions

• app.py
  Streamlit-based web application to upload an image, load the trained model (model_15.pth), and classify it as Real or Fake with probability.
• train.ipynb
  Jupyter notebook used to train or fine-tune the VGG16 model on the dataset. After VGG-based feature extraction, an SVM classifier is applied for final classification.
• model_15.pth
  Saved PyTorch model weights after training. This file is loaded during prediction in app.py.
• requirements.txt
  Contains the list of Python libraries required to run training and inference.
• README.md
  This documentation file containing project overview, setup instructions, and usage.

📦 Dataset

1. Download from the link above.
2. Extract and organize into the following folder structure:

data/
├── real/
│ ├── real1.jpg
│ └── ...
├── fake/
│ ├── fake1.jpg
│ └── ...

🧠 Training the Model

1. Load the dataset and apply image transformations.
2. Fine-tune the VGG16 model (pretrained on ImageNet).
3. Extract features using VGG16.
4. Train an SVM classifier using these features.
5. Save the final model to model_15.pth.

Note: You’ll need a GPU-enabled environment (e.g., Colab, local CUDA setup) for efficient training.

🚀 Running the Streamlit App

✅ Features of the App

• Upload an image file.
• Model loads the model_15.pth weights.
• Classifies the uploaded image as Real or Fake with probability.
• Displays prediction result and image preview.

🛠️ Requirements

• Python 3.8+
• PyTorch
• torchvision
• scikit-learn
• pandas
• Pillow
• numpy
• streamlit
• matplotlib (Install via requirements.txt)

📷 Demo

🤝 Acknowledgements