Most of the existing deepfake detection methods suffer from significant performance degradation in cross-domain settings due to distributional shifts between training and testing datasets, highlighting the urgent need for approaches capable of detecting previously unseen deepfakes. To address the challenge of limited model generalization, we decompose it into three key challenges: (1) tampering severity modeling, (2) unseen type augmentation, and (3) distribution alignment. In this paper, we propose APRNet, a novel framework designed to systematically tackle the above challenges through three core innovations. First, we devise an authentic-oriented contrastive learning mechanism that models representations across real face, intra-source forgery, and inter-source forgery, enabling fine-grained discriminative learning. Second, a progressive latent space noising strategy adaptively augments facial representations by injecting progressive noise, effectively suppressing extraneous perturbations while enhancing feature robustness. Finally, we propose relative entropy feature shift, which adjusts test-time feature distributions based on the KL divergence between test samples and the training data, allowing the model to better align inference with familiar feature distributions. Extensive experiments on multiple benchmark datasets demonstrate that APRNet outperforms state-of-the-art methods in both robustness and generalization, offering a promising direction for deepfake detection in cross-domain scenarios.