Poisoning Attacks on Algorithmic Fairness

Research in adversarial machine learning has shown how the performance of models can be seriously compromised by injecting even a small fraction poisoning points into training data. While effects on model accuracy such attacks have been widely studied, their potential other metrics remain to evaluated. In this work, we introduce an optimization framework for against algorithmic fairness, and develop gradient-based attack aimed at introducing classification disparities among different groups We empirically show that our is effective not only white-box setting, which attacker full access target model, but also more challenging black-box scenario are optimized substitute then transferred model. believe findings pave way towards definition entirely novel set targeting fairness scenarios, investigating vulnerabilities will help design robust algorithms countermeasures future.