Unfair Trojan: Targeted Backdoor Attacks Against Model Fairness

Machine learning models have been proven to have the ability to make accurate predictions on complex data tasks such as image and graph data. However, they are becoming increasingly vulnerable to various forms of attacks, such as backdoor and data poisoning attacks that can have adverse effects on model behavior. These attacks become more prevalent and complex in federated learning, where multiple local models contribute to a single global model communicating using only local gradients. Additionally, these models tend to make unfair predictions for certain protected features. Previously published works revolve around solving these issues both individually and jointly, typically by leveraging the model’s loss function to account for fairness or by adding perturbations to the unfair data. However, there has been little study on how the adversary can launch an attack that can control model fairness. This chapter demonstrates a novel and flexible attack, which we call Unfair Trojan, which aims to target model fairness while remaining stealthy. Using this attack, an adversary can have devastating effects against machine learning models, increasing their demographic parity, a key fairness metric, by up to 30%, without causing a significantly decreasing model accuracy. This chapter reveals the vulnerabilities of federated learning systems with regard to fairness and highlights the need for more robust defenses against such attacks. Our findings show the importance of understanding such attacks associated with fairness so that they can be mitigated. By revealing the ability of an adversary to exploit and amplify existing fairness issues, this chapter highlights the need for more comprehensive and proactive strategies to ensure fair predictions in machine learning applications.