The heterogeneity of IoT devices, resource bottlenecks, and security risks present challenges to the robustness and privacy of federated learning. The difference in the value of IoT data further weakens the robustness of the model. Additionally, attack methods such as Byzantine attacks and privacy inference attacks escalate the risks and vulnerabilities of federated learning models. To address these issues, we propose an asynchronous robust federated learning method based on data value. Firstly, an asynchronous grouping robust aggregation (AGRA) algorithm is designed and combined with a cosine anomaly filter to eliminate abnormal gradients. Simultaneously, an inter-group delay aggregation strategy oriented towards data value is proposed. Secondly, an efficient and robust aggregation protocol supporting asynchronous grouping is constructed to significantly reduce communication overhead while safeguarding data privacy. Finally, through data grading and secure storage adjustment using differential privacy, high-value data can be maximized, thereby enhancing the efficiency and performance of federated learning while protecting privacy. Results show that the algorithm can significantly improve accuracy by 0.9% to 65.8% in the presence of stragglers, while also reducing communication overhead by 4.2% to 5.4%.