To address the trustworthiness and robustness in federated learning and achieve effective secure aggregation, this paper presents a hierarchical model of trusted federation based on adaptive mutual learning (FedAML for short). The proposed model adopts a hierarchical structure consisting of outer loop models and local models, which enables data and model heterogeneity while ensuring that real parameters remain securely retained locally. Mutual learning is optimized by employing different temperature coefficients for efficient knowledge transfer and mitigating the accuracy issues that may arise from data heterogeneity. By leveraging adaptive temperature coefficient adjustment, clients may customize their participation, thereby allowing for greater flexibility and diversity of the local model. Our experimental analyses based on EMnist and Cifar10 datasets demonstrate that a judicious utilization of differential privacy noise can protect privacy without significantly affecting the expected model accuracy. Moreover, incorporating noise in the training process effectively defends against backdoor attacks and enhances the model’s robustness while maintaining the desired level of accuracy.