This repository includes the experiments of the bachlor's thesis "Quantitive Evaluation of the Expected Antagonism of Explainability and Privacy".
I show that the scikit-learn implementation of Individual Conditional Expectation can enable the privacy attacks membership inference and training data extraction. Additionally, counterfactuals from the explainer DiCE are shown to enable training data extraction if the method kd-tree is used.
The experiments use the datasets "Adult Data Set" by UCI and "Logistic regression To predict heart disease" from Kaggle.
Run command docker build -t xaidataprivacy . in the directory with the Dockerfile.
Then you can run the image mounted to your working directory with the following command: docker run -p 8888:8888 -v WORKING_DIRECTORY:/home/jovyan/work xaidataprivacy.
The working directory should be the directory of this repository.
Open the jupyter notebook at the URL specified in the command prompt.