Same Same, But Different?: Detecting AI-Generated Videos Using Knowledge Transfer

Video has emerged as a highly reliable medium for communication. However, recent advancements in video manipulation techniques including the generation of synthetic video content using diffusion models pose critical challenges to privacy protection and societal stability. This paper aims at presenting a detection network designed to detect AI-generated videos by exploiting knowledge transfer. Specifically, we pretrain a detector on extensive image datasets which are widely available and subsequently adapt the model on few video samples, thereby mitigating the extensive data requirements typically associated with training robust detectors. Empirical evaluations demonstrate that our VGG16-Bi-LSTM architecture achieves an accuracy and AUC of 99%, while additional testing on videos generated by three previously unseen AI models results in an AUC exceeding 92%.