ML Security Practical Session

Setup: install the requirements from the requirements.txt file:

pip install -r requirements.txt

First, read the tutorial notebook. It can be found here.

Then, we are going to write a pipeline for running a security evaluation experiment.

Complete the parts that are missing in the file pipeline_for_robustness_evaluation.py. They are marked with a TODO comment (automatically recognized by most IDEs).

For running the debugging script, issue the following commands in the terminal:

cd src
python -m pipeline_for_robustness_evaluation --help

First, debug the adversarial attack, find out if the attack is working well with the configuration you defined.

python -m pipeline_for_robustness_evaluation --model 0 --debug

Once everythin is fine, we can finally run the complete evaluation on more samples (recommended GPU, or limit the number of samples to a small value).

cd src
python -m pipeline_for_robustness_evaluation --model 0 --samples 5