Deepfake detectors nowadays have achieved impressive performance on intra-dataset evaluation, where training and testing images are collected from the same dataset. However, their detection results decrease when dealing with deepfake samples from unknown data collection. This paper introduces a novel Pixel-level Face Correction (PFC) task that involves pre-training with authentic facial data. Specifically, PFC generates pixel-level discrepancy by constructing a multi-scale facial pyramid and applying pixel-level resolution blending technique, then leverages masked image modeling framework to correct the subtle difference and learn facial representations that can be transferred to unseen deepfake detections. Extensive experimental results show that the proposed method obtains robust facial reconstruction ability and improves the model generalization to unknown manipulations.