<p>To counter deepfakes, significant progress has been made in deepfake detection in recent years. However, existing methods still face generalization issues dealing with various forgery types, which limits their applicability in real-world scenarios. This study reveals that the focus regions of the detector differ when detecting images generated by different forgery methods, which significantly affects its generalization performance. To address this issue and improve generalization ability, this paper proposes a novel detection framework that incorporates data interaction and feature interaction strategies. First, an inter-class data interaction module is proposed to facilitate information interaction between real and fake face images. By embedding characteristics of real samples into fake ones, it effectively reduces the sensitivity of the model to specific forgery traces, guiding it to focus on common discriminative regions. Then, leveraging the complementarity between frequency and spatial-domain features, a dual-domain feature interaction module is designed. This module employs a non-linear exponential scaling strategy to adaptively integrate key information from both domains, effectively enlarging the inter-class feature distances. Finally, a cosine metric center loss is introduced, aiming to maximize intra-class similarity and inter-class separability, thus improving the model generalization performance. Experimental results demonstrate that the proposed method significantly enhances the generalization ability of the detector and outperforms some state-of-the-art approaches across multiple benchmark datasets.</p>