Federated learning (FL) enables multiple participants to collaboratively train a shared global model by uploading locally trained updates while preserving data privacy. However, its distributed nature and privacy-preserving mechanisms make FL vulnerable to backdoor attacks, which have become one of the most critical security threats. To achieve greater stealth and persistence, adversaries often inject backdoors into gradient directions irrelevant to the main task and leverage large updates to task-sensitive parameters associated with benign tasks to mask subtle backdoor manipulations. This blurs the boundary between benign and malicious updates, reducing the effectiveness of existing detection methods. To address this challenge, we begin by defining dynamic dimensions as the model dimensions that exhibit consistent changes across multiple training rounds. Based on this, we propose StealthFL, which identifies dynamic dimensions through analyzing deviations from historical updates, thereby reducing the masking effect of task-sensitive parameters. It then applies discrete wavelet transform (DWT) to filter suspicious backdoor components within these dynamic dimensions and employs Mahalanobis distance for anomaly detection. Extensive experiments under various system configurations and attack scenarios demonstrate that StealthFL outperforms existing defenses in both effectiveness and robustness.