detector_template

Detector template

This directory includes the basic components used to incorporate a new detector into the system. In general terms the procedure is as follows:

• Add the implementation for your specific detector and wrap it in the ensemble class.
• Build and test the docker image and container.
• Add the container to the docker-compose.yml file.

The following section covers these steps in more detail.

Adding a new detector

1. Create a new detector folder by copying FakeFinder/detector/detector_template to FakeFinder/detectors/<your_detector>. You will be working inside this directory:

cd detectors/<your_detector>

<your_detector> should be a short, single-word name for your detector (e.g. selimsef or ntech).

2. Implement the new detector in this directory. Wrap it into an ensemble class using Ensemble.py. The basic ensemble class includes two methods:
   1. The __init__ method should initialize the model and load any necessary weights from the local weights directory. Ideally this ensemble class should load all necessary weights, but this can occur in step 3 if so desired.
   2. The inference method will be called to make the predictions. This method should expect to receive the path to a video, run the forensics pipeline for that algorithm, and return the score. Currently we assigned the following labels:
      • 1: fake/manipualted
      • 0: real.
3. Make any necesary changes to app.py. This file will be used to initialize a flask application used by the service to respond to requests.
4. Generate a requirements file “requirements.txt” listing the packages and version required to recreate the environment used for the new detector. For example you may use the pipreqs package:

pip install pipreqs
pipreqs .

5. Edit the docker file with any additional steps needed to reproduce your working environment. The template file we have included will:
   • Create a basic linux + python3 system
   • Install the packages in the requirements file
   • Expose port 5000 for external access to/from the API
   • Run app.py to start up the flask application You can test the dockerfile build by replacing the last line:

CMD ["python3","app.py"]

with

CMD ["/bin/bash"]

This will change the state of the docker container to provide a shell for experimentation instead of running the flask app. Once the container has been tested return to the original:

CMD ["python3","app.py"]

You can also override the default python3 app.py command by running the container with the following command:

docker run -it --entrypoint=/bin/bash <your_image_name> -i

Test to make sure that an instance of the ensemble class performs inference without error, and that running app.py initializes an instance of ensemble withour error.

6. Push the code to your remote repository for use later. Make sure that your weights are available in the ./weights/<your_detector>/ directory so that it can be correctly mounted when running the container.