Human vs. Automatic Detection of Deepfake Videos Over Noisy Channels

Identification of DeepFake video content is a challenging scientific problem that addresses a growing societal concern. We investigate the relationship between DeepFake detection by humans and by automatic methods based on state-of-the-art deep learning algorithms. The main novelty of our work is the consideration of videos that are transmitted through noisy channels and arrive with distortions. This reflects many practical environments, including surveillance based on cameras connected via noisy wireless links and videoconferencing in driving vehicles. We conduct a user study with 192 probands who classify real (genuine) and DeepFake videos with and without various classes of distortions. We find that today's deep neural networks (DNNs) outperform humans by far, whereas humans are heavily distracted by random noise from the channel. Moreover, DNNs are robust under distortions, achieving perfect classification on distorted data even when trained on distortion-free content. It appears that the human visual system and DNNs are approaching the DeepFake classification problem quite differently and their respective strengths and weaknesses are largely uncorrelated.