The paper "DF40: Toward Next-Generation Deepfake Detection" proposes a new comprehensive benchmark, DF40, to revolutionize the deepfake detection field. The authors identify that existing deepfake detection datasets and evaluation protocols are limited in their diversity and realism, leading to a gap between the performance of detectors on these datasets and their effectiveness in real-world scenarios. DF40 addresses this issue by constructing a highly diverse and large-scale dataset that includes 40 distinct deepfake techniques, encompassing face-swapping, face-reenactment, entire face synthesis, and face editing. The dataset is designed to simulate realistic and diverse deepfakes, including methods that are not commonly used in existing datasets. The paper conducts extensive evaluations using four standard evaluation protocols and eight representative detection methods, resulting in over 2,000 evaluations. The findings from these evaluations provide new insights into the performance of deepfake detectors and highlight several limitations of current methods. Specifically, the study reveals that existing state-of-the-art (SoTA) detectors may not have significant advantages over simpler baselines, and that certain deepfake types, such as face-reenactment, may share transferable patterns that can be leveraged for detection. The paper also discusses the importance of a diverse dataset in improving the generalization of deepfake detectors and explores the impact of super-resolution images and frequency artifacts on detection performance. The authors conclude that DF40 offers high-quality and realistic deepfake techniques, facilitating the detection of today's real-world deepfakes and contributing to societal trust and responsible technology use. They also outline several open questions and potential topics for future research, including the role of blending data, the diversity of deepfakes, and the development of domain-invariant detectors.The paper "DF40: Toward Next-Generation Deepfake Detection" proposes a new comprehensive benchmark, DF40, to revolutionize the deepfake detection field. The authors identify that existing deepfake detection datasets and evaluation protocols are limited in their diversity and realism, leading to a gap between the performance of detectors on these datasets and their effectiveness in real-world scenarios. DF40 addresses this issue by constructing a highly diverse and large-scale dataset that includes 40 distinct deepfake techniques, encompassing face-swapping, face-reenactment, entire face synthesis, and face editing. The dataset is designed to simulate realistic and diverse deepfakes, including methods that are not commonly used in existing datasets. The paper conducts extensive evaluations using four standard evaluation protocols and eight representative detection methods, resulting in over 2,000 evaluations. The findings from these evaluations provide new insights into the performance of deepfake detectors and highlight several limitations of current methods. Specifically, the study reveals that existing state-of-the-art (SoTA) detectors may not have significant advantages over simpler baselines, and that certain deepfake types, such as face-reenactment, may share transferable patterns that can be leveraged for detection. The paper also discusses the importance of a diverse dataset in improving the generalization of deepfake detectors and explores the impact of super-resolution images and frequency artifacts on detection performance. The authors conclude that DF40 offers high-quality and realistic deepfake techniques, facilitating the detection of today's real-world deepfakes and contributing to societal trust and responsible technology use. They also outline several open questions and potential topics for future research, including the role of blending data, the diversity of deepfakes, and the development of domain-invariant detectors.