Deepfake in video conferencing

Introduction

This article explains how one can project a deepfake-ed live video stream of themselves to a meeting on Skype, Zoom, GoogleMeet or any other video calling platform using python, opencv and deepfacelive app running in a docker container. Source code/sample scripts used in this article can be found in this repository.

Solution Overview

Instructions

• Start the DeepFaceLive app and make it output a UDP stream
• Read the UDP stream and write it to a virtual cam

Start DeepFaceLive app

1. Clone the deepfacelive repo

    git clone https://github.com/iperov/DeepFaceLive
   

2. Add the ‘–network host’ flag to docker run in ‘start.sh’ script to let the container use host network

    cd DeepFaceLive/build/linux/
    nano start.sh
   
    docker run --network host --ipc host --gpus all -e DISPLAY=$DISPLAY -v /tmp/.X11-unix:/tmp/.X11-unix -v $DATA_FOLDER:/data/ $CAM0 $CAM1 $CAM2 $CAM3  --rm -it deepfacelive
   

3. Build and start the image for Linux, make sure to pass the ‘-c’ flag to pass host machine cameras to docker

    ./start.sh -c
   

4. Once the app has started, select the camera by choosing a device number from the dropdown. Usually, your system camera is at 0th index, if it is not try a different number. Once successful, your camera stream should be displayed in the window below.

   

5. Load a pre-trained face model of a character that you want to swap your face with. Note downloading the model may take some time.

   

6. Once you can see a swapped result, selected ‘Merged output’ from the drpodown and tick the checkbox to output result as a UDP stream.

   

   Note the udp address 127.0.0.1:1234, we will use this to read the mpeg transport stream in opencv.

7. To verify that output stream is being written to the defined port, open a terminal window and run the following command.

    tcpdump -i lo -n udp port 1234
   

   The output should look something like below

Start pyvirtualcam

1. Clone deepfacelive-virtualcam script:

    git clone https://github.com/marib-suItan/deepfacelive-virtualcam
    cd deepfacelive-virtualcam/
   

2. Create a virtual camera using v4l2loopback and verify its existence.

    sudo apt install v4l2loopback-dkms
    modprobe v4l2loopback video_nr=2
    ls -1 /sys/devices/virtual/video4linux
   

3. Run the python script

    python virtualcam.py
   

4. If everything has been successful you should see a dummy cam option alongside your system camera in other applications.

   Note Chrome seems to block virtual cameras for some reason, but Firefox, Skype and MSTeams work just fine.