Asynchronous Federated Learning (AFL) is a distributed learning technique that, building on the privacy protection provided by Synchronous Federated Learning (SFL), allows nodes to train and upload models locally at their own pace. This approach addresses the issue of slow devices that hinder SFL’s performance. However, data and device heterogeneity present significant challenges to the training effectiveness of AFL. In this paper, we conduct a detailed analysis of the factors that impact the training performance of AFL in the context of typical neural network classification tasks in federated learning. Taking into account both data and device heterogeneity, as well as privacy protection requirements, we propose Federated Asynchronous Heterogeneity-Adaptive Learning with Output Estimation (FAHALE). By leveraging the client aggregation regulation and global model aggregation weighting mechanisms we introduced, our method achieves up to 3x faster convergence compared to several state-of-the-art methods, while demonstrating significant advantages in training stability and achieving a collaborative optimization of privacy, utility, and security.