This repository contains the code for conducted experiment regarding the master thesis "A systematic impact analysis of common post-training pre-processing methods to mitigate racial bias in face verification". The structure is the following:

• evaluation_notebooks contains all notebooks used to conduct the analysis of the experiment results. Each processing method has its own notebook witht he results for all chosen models.
• experimentation is the folder which contains one notebook per processing method to execute the experiment
• results contains all csv with the data used for experiments and their corresponding results. These are later used in notebooks to evaluate the results

To spin up a Docker container to run these experiments please go to the root folder of this repository and run docker-compose up

The notebooks can be exectued by running all cells. This should work without any error message.

Before running the notebooks please download the file Equalizedface.tar.gz from this link:

https://console.cloud.google.com/storage/browser/racial_bias_data;tab=permissions?forceOnBucketsSortingFiltering=false&hl=de&project=racialbias&prefix=&forceOnObjectsSortingFiltering=false

The following emails have currently access to the data:

• [email protected]
• [email protected]

After downloading the file please unzip the file and put the contained folder named race_per_7000 into a folder called data/ in the root directory.

To re-create the experimentation data execute all notebooks in experimentation. However this should not be needed as all experimentation data has been created already and lives under data/val_data.

The evaluation notebooks can be executed and and access the stored results.