Online document presentation attack detection (DPAD) in the wild is now advancing towards generalization and robustness. This paper proposes a teacher-student framework with a robust student network capable of handling high-quality and low-quality recaptured images in real-world scenarios when the training set is limited to high-quality data. Our framework addresses the challenge by utilizing diffusion-based simulation methods and knowledge distillation between branch networks. Firstly, we employ the low-quality distortions simulation (LDS) module to synthetic degrade images corresponding to high-quality images, serving as input data for the student network. Then, cross-quality distillation (CQD) is applied at the feature layer before the classification head to constrain the student network to learn rich pre-training high-quality features from the teacher network. Combining the LDS and CQD modules, the student network unifies the forensic features of recaptured images with various qualities and can robustly detect them in practical settings. Experimental results demonstrate the effectiveness of our proposed detection method on the DLC2021, RSCID (N), and RSCID (L) datasets, outperforming six large-scale detection backbones when using only ResNet18 as the student network.