Malignant brain tumors pose a significant global threat, emphasizing the critical need for efficient diagnostic methods utilizing MRI. Manual analysis of MRI images is labor-intensive and subjective, highlighting the necessity for faster and automated effective methods. In this paper, we propose an uncertainty-aware robust information fusion attention network model for precisely classifying brain tumors in MRI images. Our approach introduces a novel robust information fusion attention layer that learns enhanced representations by integrating global context with local information. We estimate the uncertainty in our model’s predictions using the ensemble Monte Carlo dropout strategy. Our findings demonstrate outstanding performance, achieving accuracies of 98.37% on the Cheng dataset and 98.48% on the Nickparvar dataset in brain tumor MRI image classification tasks, while minimizing computational costs in terms of resource usage and inference time.