In federated learning

the issue of noisy labels severely limits the practical performance of models. In real-world scenarios

the quality of client data often varies significantly: some clients have relatively clean labels

while others contain a large amount of noise. Existing methods typically assume that clients follow a uniform noise distribution

making it difficult to effectively handle the complex situation of mixed-noise clients. To address this

we propose a novel federated learning framework (FedCDA) that achieves robust learning for groups of mixed-noise clients through a multi-stage design. First

the framework uses self-supervised contrastive learning for unsupervised client noise assessment

thereby dividing clients into clean and abnormal (high-noise) groups. Second

a differentiated training strategy is employed to assign appropriate linear classification weights to clients in different groups and train a robust classifier. Finally

a teacher-adaptive knowledge distillation mechanism is designed to dynamically distill and adjust client local models

ensuring the reliability of the training process. Extensive experiments show that our proposed framework achieves superior experimental results in federated learning scenarios with mixed noise compared to existing methods.