This paper proposes a novel deepfake detection framework that aims to extract broader forgery clues by combining local and global information losses. The framework addresses the limitations of existing deepfake detection methods, which often suffer from overfitting and lack theoretical guarantees for extracting sufficient forgery clues. The proposed method introduces two key objectives: Local Information Loss (LIL) and Global Information Loss (GIL). LIL ensures that local features are orthogonal and contain comprehensive task-relevant information, while GIL aggregates these features into a global representation that retains sufficient task-related information and eliminates superfluous information. The method is evaluated on five benchmark datasets, including FaceForensics++, Celeb-DF, and DFDC, and achieves state-of-the-art performance in both in-dataset and cross-dataset settings. The results demonstrate that the proposed method is more robust and generalizable compared to existing approaches, with significant improvements in detection accuracy across different datasets. The method also provides insights into the types of forgery clues that can be effectively detected, highlighting the importance of considering multiple non-overlapping regions in deepfake detection. The framework is designed to be theoretically grounded and computationally efficient, making it a promising solution for real-world deepfake detection applications.This paper proposes a novel deepfake detection framework that aims to extract broader forgery clues by combining local and global information losses. The framework addresses the limitations of existing deepfake detection methods, which often suffer from overfitting and lack theoretical guarantees for extracting sufficient forgery clues. The proposed method introduces two key objectives: Local Information Loss (LIL) and Global Information Loss (GIL). LIL ensures that local features are orthogonal and contain comprehensive task-relevant information, while GIL aggregates these features into a global representation that retains sufficient task-related information and eliminates superfluous information. The method is evaluated on five benchmark datasets, including FaceForensics++, Celeb-DF, and DFDC, and achieves state-of-the-art performance in both in-dataset and cross-dataset settings. The results demonstrate that the proposed method is more robust and generalizable compared to existing approaches, with significant improvements in detection accuracy across different datasets. The method also provides insights into the types of forgery clues that can be effectively detected, highlighting the importance of considering multiple non-overlapping regions in deepfake detection. The framework is designed to be theoretically grounded and computationally efficient, making it a promising solution for real-world deepfake detection applications.