Current deepfake detection methods focus on learning specific forged traces, but they struggle with unknown forgery types. To address this issue, we propose a noise feature consistency-based approach. We utilize both spatial and noise features in face images, as noise features effectively capture forged traces. To achieve robust feature representations, we design a cross-attention module to interact between noise and spatial features. Additionally, we design a comprehensive consistency guidance module to consider both intra- and inter-instance feature consistency. Experiments prove that our proposed method has good robustness and generalization.