A Multi-Stream Fusion Approach with One-Class Learning for Audio-Visual Deepfake Detection

---

This repository contains the official implementation of our paper, "A Multi-Stream Fusion Approach with One-Class Learning for Audio-Visual Deepfake Detection"

Requirement

---

python==3.9 \ pytorch-lightning=1.7.7

Data Preparation

---

Need to create a new csv file for our train, validation, and test datasets by running add_meta_data.ipynb Need to run create_pairs.ipynb and voice_converion.py to create voice conversion fake videos which is part of our test set. Need to update path variable before running create_pairs.ipynb.

Run the training code

---

Need to change DATA_ROOT and OUTPUT before running the scripts/run.sh

Evaluation

---

Need to change DATA_ROOT and OUTPUT before running the scripts/eval_all.sh