Recent advances in deepfake technology have raised significant concerns regarding identity manipulation, driving rapid progress in detection methods. While traditional low-level feature-based detection methods struggle with robustness in real-world scenarios, high-level feature-based methods offer improved robustness. However, they are prone to overfitting, limiting their performance on unseen datasets. To address these challenges, we propose the Masked Identity Consistency Detector (MICD), a novel framework that detects forgeries by leveraging the consistency between internal and external identity. We introduce an adaptive masking strategy to optimize the model’s receptive field based on the attention map, enhancing generalization while decoupling internal and external identity features. Additionally, we propose two specialized loss functions: Inconsistency loss, which amplifies the difference between internal and external identity features, and Redundancy loss, which minimizes redundant information during feature fusion. Experimental results show that MICD achieves superior performance compared to state-of-the-art methods on commonly used datasets, demonstrating its effectiveness in both in-dataset and cross-dataset scenarios.