Detection and Attribution of Diffusion Model of Character Animation Based on Spatio-Temporal Attention

Fazhong Liu, Yan Meng, Tian Dong, Guoxing Chen, Haojin Zhu

Abstract: Character animation technology has made significant progress in recent years, enabling the generation of highly realistic videos. However, this technology can also be used to create fake videos that are difficult to distinguish from real ones, posing a threat to social media and online platforms. In this study, we propose a novel detection and attribution method that uses spatial-temporal attention mechanisms to identify features of model outputs. We evaluated our method in dataset containing 2,924 video samples from TikTok and four video generation models. Our results show that our proposed method can correlate the generated video with the real model with 73.68% accuracy, outperforming existing methods. This study provides a foundation for users and researchers exploring the field of video generation and fake video detection.

External IDs:doi:10.1145/3689217.3690623