Ethical Adversaries

Machine learning is being integrated into a growing number of critical systems with far-reaching impacts on society. Unexpected behaviour and unfair decision processes are coming under increasing scrutiny due to this widespread use its theoretical considerations. Individuals, as well organisations, notice, test, criticize results hold model designers deployers accountable. We offer framework that assists these groups in mitigating representations stemming from the training datasets. Our relies two inter-operating adversaries improve fairness. First, trained goal preventing guessing protected attributes' values while limiting utility losses. This first step optimizes model's parameters for Second, leverages evasion attacks adversarial machine generate new examples will be misclassified. These then used retrain step. steps iteratively applied until significant improvement fairness obtained. evaluated our well-studied datasets literature - including COMPAS where it can surpass other approaches concerning demographic parity, equality opportunity also utility. investigated trade-offs between targets terms hyperparameters illustrated findings subtle difficulties when unfairness highlight how assist designers.