Open domain adaptation is a significant area of research that transcends the limitations of closed domain adaptation. Although numerous effective methods have been developed in recent years, it remains essential to address challenges such as data security and noisy labels within the source domain dataset, as these issues frequently arise in real-world applications. To tackle these concerns, we propose Noise-Resistant Federated Learning (NRFed). First, we integrate federated learning to ensure data security by keeping the data on local clients. Next, we incorporate Bayesian Neural Networks (BNNs), which enable local clients to assess their own uncertainty and achieve weighted aggregation, thereby reducing the impact of noisy clients. To further address the issue of noisy labels, we implement a confidence-based label correction technique that purifies the noisy training dataset. Finally, extensive experiments conducted on the Office-31 and Office-Home datasets demonstrate that NRFed effectively mitigates issues related to noisy labels.