Multi-step and Iterative Backdoor Injection in Federated Machine Unlearning

This paper investigates multistep and iterative back-door injection in federated learning, revealing how gradual and distributed attacks enhance the persistence and stealth of backdoors. These attacks evade detection by standard defenses, complicating the process of machine unlearning. Experimental evaluations show that traditional unlearning methods are only partially effective, while our proposed iterative unlearning strategy significantly reduces backdoor success without degrading model accuracy. The findings emphasize the need for adaptive and robust unlearning techniques designed for the unique challenges of federated learning, offering new directions for defending against sophisticated adversarial attacks.