seminar_fairness - Evaluation of Fairness Frameworks for Robustness against Adversarial Attacks

This repository implements and evaluates multiple fairness frameworks applied to NLP text classification under adversarial settings. The implementations include:

• FRAPPÉ on NLP Text Classification: A framework that applies post-processing fairness mitigation.
• BadFair on NLP Text Classification: An approach to attack fairness mechanisms via data poisoning.
• BadFair on FRAPPÉ: Integration and cross-evaluation of the BadFair attack within the FRAPPÉ framework.

Usage guide

Prerequisites

Python pacakge requirements are in requirements.txt and need to be installed

Workflow

There are 4 folders, 2 for each dataset, 1 with the Badfair attack and 1 without. Since training data poisoning and test data manipulation is needed to implement Badfair, this folder seperation was done to prevent accidental cross-contamination. The logits are saved in the contents folder which needs to be emptied every time parameters are changed and the model is retrained.

In each folder train_model.py should be run first and then evaluate.py.

• train_model.py downloads (and poisons the data) and trains the model and applies the FRAPPE framework.
• evaluate.py calculates Fairness metrics and calls postproc_fairness/fairmain.py
• postproc_fairness/fairmain.py applies FRAPPE and calculates Fairness metrics