<p>In the smart cockpit, the accuracy of facial recognition is reduced heavily under low-light conditions. Due to the insufficient lighting, this limitation poses challenges for smart cockpits in obtaining features in dark environments. To address this challenge, we propose a multi-task learning model utilizing hierarchical feature disentanglement that processes features at varying granularities to optimize facial recognition. Our model includes tasks for gender classification, identity prediction, and image reconstruction, respectively. Specifically, the gender classification task enhances the model’s ability to identify coarse-grained facial features, thus aiding in the accurate initialization of more detailed identity prediction tasks. Concurrently, the identity prediction task leverages finer-grained features to improve personalized recognition accuracy under variable lighting conditions. The image reconstruction task utilizes multi-scale feature representations to deepen the model’s comprehension of facial features for enhancing overall recognition robustness. We develop a specialized dataset to simulate low-exposure scenarios in smart cockpits, facilitating realistic training conditions. Our experimental results demonstrate that this structured multi-task approach significantly enhances the model’s performance in low-exposure environments, by effectively utilizing features across different granularities. The dataset, along with the code, is publicly available in a repository.</p>