Recently, deep neural networks have achieved significant success in face recognition tasks. However, direct application to low-resolution (LR) face recognition results in noticeable degradation due to the lack of details. Existing methods primarily distill prior knowledge from high-resolution (HR) face network to facilitate LR face recognition but lack the ability to transfer multi-level knowledge. To address this issue, we propose a collaborative knowledge distillation approach, which transfers both attention maps and logits obtained from the HR network as a teacher to LR network as a student to enhance LR recognition performance. The attention maps and logits represent knowledge from the intermediate layers and the output, respectively. We design the knowledge distillation loss using the maximum mean discrepancy as the distance measure between the teacher and the student networks’ attention distributions. Furthermore, the attention mechanism assigns different degrees of importance to facial regions. Extensive experiments on face recognition benchmarks demonstrate the effectiveness of our approach. Our method outperforms the baseline model and other state-of-the-art knowledge distillation methods in transferring knowledge across various resolutions.