deepfake-detector

FINAL YEAR PROJECT ON DEEPFAKE DETECTION WITH CNN (USING PYTORCH)

Abstract

The rise of deepfakes presents a significant threat to digital media authenticity and trust. This project presents a deepfake detection pipeline based on InceptionResNetV1, pre-trained on VGGFace2 and fine-tuned on a balanced dataset of 206,348 preprocessed facial images extracted from the Celeb-DF, Faceforensics and Openforensics datasets with a stratified 70/30 split ensuring fair training and validation distribution. Despite being trained on CPU for only 20 epochs, the model achieved a final validation accuracy of 97.15% and a validation loss of 0.0671, indicating strong generalization and low overfitting. The study highlights the viability of using computationally efficient training setups alongside proper data preparation for high-accuracy deepfake detection. Future work will explore video-level analysis, ensemble learning and real-time deployment strategies to expand the model’s applicability.

Directory Structure

Recommended Directory Structure

Model Architecture

Tools and Technologies

1. Python, Programming Language.
2. PyTorch, ML Framework.
3. Datasets: Ensemble Dataset of FaceForensics, OpenForensics and Celeb-DF V2.
4. Visualization: Matplotlib.
5. Deployment: Streamlit App.

Author

Charles Ekanem
B.Eng Student, Computer Engineering
LinkedIn
Twitter
Website

Project setup

git clone https://github.com/Charles04Ekanem/deepfake-detector.git
cd deepfake-detector
pip install -r requirements.txt