Federated Learning (FL), as a distributed model training paradigm, has garnered significant attention and practical application. Recently, increasing research efforts have focused on Personalized Federated Learning (PFL) as an effective solution to address data heterogeneity in FL systems. However, existing studies reveal that PFL remains vulnerable to stealthy yet harmful backdoor attacks. Furthermore, current federated learning (FL) algorithms designed to defend against backdoor attacks demonstrate significant degradation in model performance and substantial decline in defensive effectiveness when applied to personalized scenarios. To bridge this research gap, we propose a robust PFL framework against backdoor attacks. Our framework incorporates a three-tier defense mechanism: (1) Clients initially purify potential model backdoors through adversarial example generation; (2) An alternating training strategy for hierarchical models is employed to block backdoor attacks while generating personalized head models; (3) The server implements a trimmed aggregation mechanism to mitigate malicious client impacts. Comprehensive experiments on three benchmark datasets demonstrate the framework’s effectiveness, showing superior model performance in personalized scenarios while achieving up to 89% attack success suppression rate compared with eight state-of-the-art defense methods.