SHIELD: A Specialized Dataset for Hybrid Blind Forensics of World Leaders

The speech videos of public figures, such as movie celebrities and world leaders, have an extensive influence on the Internet. However, the authenticity of these videos is often difficult to ascertain. These videos may have been carefully imitated by comedians or manipulated using Deepfake methods, which creates significant obstacles for the video forensics of specific characters. Moreover, the vast amount of data on social networking platforms renders manual screening impractical. To specifically address this issue, we present SHIELD, which stands for Specialized dataset for Hybrid blInd forEnsics of worLd leaDers. Unlike most previous public Deepfake datasets that only contain Deepfake samples, this dataset exquisitely includes a collection that can quickly test this issue, encompassing both impersonator and Deepfake videos. We provide a detailed dataset production process and conduct an elaborate experiment under the hybrid blind detection scenario. Our findings reveal the limitations of existing methods, demonstrate the potential of identity-based models, and illustrate the increased challenges posed by SHIELD.