TRUSTED.AI

Artificial intelligence (AI) systems are getting more and more relevance as support to human decision-making processes. While AI holds the promise of delivering valuable insights into many application scenarios, the broad adoption of AI systems will rely on the ability to trust their decisions. This python application enables developers to automatically compute the thrustworithness level for their machine learning models. We support the most common machine learning libraries like TensorFlow, Sklearn and PyTorch.

1. Time Schedule

• Analyze the current state of the art regarding thrustworithniess. (Februar - March)
• Identify missing aspects relevant for automatically calculating a model's trust score (April)
• Create a taxonomy containing relevant pillars and metrics (May)
• Select and analyze a suitable application scenario (June)
• Design a trusted AI algorithm (July)
• Implementation (August - September)
• Evaluation and Discussion (October)
• Documentation and Process (November)

2. Deployment

2.1 Local Deployment

The webapp is build using Flask, Dash and is optimized to run on python version 3.9.7. In order to setup and run the webapplication you need to execute the following steps. At first open the console and navigate to the webapp directory. Then install the necessary dependencies form the requirements.txt file. Start the application.

cd webapp
pip install -r requirements.txt
python index.py

2.2 Containerized Deployment

For easy deployment, our project can be combined into a Docker container, using the following commands.

1. Run docker build command in order to build the Docker Image from a Dockerfile

> docker build -f ./webapp/Dockerfile -t trustedai/webapp:v1 .
# docker build -f <path-to-dockerfile> -t <hub-user>/<repo-name>:<tag> .
# -f ... specify path to Dockerfile
# -t ... add a tag to the image

2. Login to Docker Hub

> docker login --username trustedai
# docker login --username <hub-user>

3. Push the Docker image to Docker Hub

> docker push trustedai/webapp:v1
# docker push <hub-user>/<repo-name>:<tag>

4. Pull the Docker image from Docker Hub on the remote server

> docker pull trustedai/webapp:v1
# docker pull <hub-user>/<repo-name>:<tag>

5. Run the Docker image as a local Docker container and check if everything is working correctly

> docker run -p 80:8080 trustedai/webapp:v1
# docker run -p <host-port>:<container-port> <hub-user>/<repo-name>:<tag>

3. Documentation

The whole documentation of the project can be found on our GitHub page The documentation is created with mkdocs, which uses simple markdown file and a .yml config file to build and deploy the documentation to GitHub.

The relevant files to create the documentation are in the mkdocs folder

The file mkdocs.yml is the config file and has to be adapted if the theme should change or the index changes or new markdowns added. The markdown files are in the folder docs and include the whole documentation.

installation

pip install --upgrade pip
pip install mkdocs
pip install pymdown-extensions
pip install python-markdown-math

workflow:

1. At the location of the mkdocs folder open the terminal and enter:

   mkdocs serve
   

   This will build the documentation and host it locally, where the full documentation can be seen.

2. change the config or markdown files. Sinc you are serving the documentation locally you can see the changes in the config or markdown files in real time. You will always see the output and how the documentation will look in the end and therefore minimalize syntax errors in markdown for example.

3. deploy the build on GitHub to the branche gh-pages. For this open another termianl at the localtion of the mkdocs folder and enter:

   mkdocs gh-deploy
   

4. commit and push the changed config or markdown files (keep the source of the build up to date)