The prevalence of label noise in datasets poses significant challenges in supervised learning frameworks, where models that overfit noisy labels experience degraded generalization performance. Numerous robust learning methodologies have been proposed to mitigate the adverse effects of noisy labels on models. Among these methodologies, sample selection-based approaches have garnered notable attention for their promising results on real-world noisy datasets. However, current research predominantly aims at improving the overall accuracy of models in the presence of label noise, often overlooking the critical aspect of fairness across different classes. In this paper, we argue that ensuring model fairness is as crucial as maintaining robustness in the context of label noise. We propose a novel approach that combines advanced sample selection and adversarial optimization to enhance both fairness and robustness simultaneously. Our methodology introduces implicit regularization to model label noise and proposes a sample selection strategy based on the distribution of noise probabilities and associated loss values. Furthermore, we decouple representation learning from classification head learning by leveraging adversarial optimization, focusing on the gradients of the worst-case classification hyperplane. Experimental comparisons on both synthetic and real-world noisy datasets demonstrate that our proposed method achieves superior performance and optimal class fairness. The effectiveness of our approach is substantiated by empirical results, and we provide comprehensive evaluations detailing the robustness against label noise. The code implementing our methodology will be made publicly available at https://github.com/wangnaihao/DNLL.git , facilitating further research and development in the field.