DF40: Toward Next-Generation Deepfake Detection This paper introduces DF40, a comprehensive deepfake detection benchmark that addresses the limitations of existing deepfake detection methods. Current methods often rely on specific datasets, leading to limited generalization to real-world deepfakes. DF40 contains 40 distinct deepfake techniques, including face-swapping, face-reenactment, entire face synthesis, and face editing, making it significantly more diverse and extensive than existing datasets. The dataset includes over 40 deepfake techniques, with 10 times more techniques than FF++. The paper evaluates various detection methods using four standard protocols and eight representative methods, resulting in over 2,000 evaluations. The results reveal seven new insights, including the importance of fake region in detection, the limitations of existing SoTA detectors, and the effectiveness of CLIP in deepfake detection. The paper also highlights the need for more diverse deepfake datasets to improve detection performance. The authors also identify four open research questions for future work, including the role of blending data in training deepfake detectors, the design of new frameworks for learning diverse forgeries, the classification of deepfakes based on FS, FR, EFS, and FE, and the development of domain-invariant detectors. DF40 provides a comprehensive benchmark for deepfake detection, offering a diverse set of deepfake techniques and evaluations under four standard protocols. The results show that existing detectors may not perform well on new deepfake types, highlighting the need for more diverse datasets. The paper also emphasizes the importance of pre-training in deepfake detection and the potential of CLIP in this area. The authors hope that DF40 will revolutionize the field of deepfake detection for the next generation.DF40: Toward Next-Generation Deepfake Detection This paper introduces DF40, a comprehensive deepfake detection benchmark that addresses the limitations of existing deepfake detection methods. Current methods often rely on specific datasets, leading to limited generalization to real-world deepfakes. DF40 contains 40 distinct deepfake techniques, including face-swapping, face-reenactment, entire face synthesis, and face editing, making it significantly more diverse and extensive than existing datasets. The dataset includes over 40 deepfake techniques, with 10 times more techniques than FF++. The paper evaluates various detection methods using four standard protocols and eight representative methods, resulting in over 2,000 evaluations. The results reveal seven new insights, including the importance of fake region in detection, the limitations of existing SoTA detectors, and the effectiveness of CLIP in deepfake detection. The paper also highlights the need for more diverse deepfake datasets to improve detection performance. The authors also identify four open research questions for future work, including the role of blending data in training deepfake detectors, the design of new frameworks for learning diverse forgeries, the classification of deepfakes based on FS, FR, EFS, and FE, and the development of domain-invariant detectors. DF40 provides a comprehensive benchmark for deepfake detection, offering a diverse set of deepfake techniques and evaluations under four standard protocols. The results show that existing detectors may not perform well on new deepfake types, highlighting the need for more diverse datasets. The paper also emphasizes the importance of pre-training in deepfake detection and the potential of CLIP in this area. The authors hope that DF40 will revolutionize the field of deepfake detection for the next generation.