Deepfake Detection with Deep Learning

Marten Thompson

Overview

This repo contains all the input data, code, and output required for our research. If you wish to execute code, the python scripts are likely the easiest to implement; Colab notebooks require a copy of the original data in your Google Drive.

Contents

• code all code required to perform research.
  • The script analysis.py and notebook analysis.ipynb both serve as the primary point of interaction and execution for training models, the latter for Colab. The notebook transfer.ipynb similarly manages working with pre-trained models.
  • data_mgmt.py manages data on disk. The first time you train a model, this script will organize the appropriate test/train data organization. In the case of 3D data, it writes 100Gb of numpy arrays.
  • models.py and models3D.py contain functions that when called create, train, and save models. They also contain a wrapper class definition for convenience.
• data/original untouched DFDC data. Functions in data_mgmt.py will write to the data directory when organizing your local environment.
• report written report, slides, and recording of presentation.
• saved_models all the final models, saved in tensorflow format
• small_local point functions here for small local testing
• visualization
  • visuals.ipynb script for making training diagram
  • model_diagrams images of different network architectures
  • plotting_data all the train/test logs used for graphics