Zsolnai breaks down how deepfake detectors work and why they matter. Face2Face and FaceForensics datasets show that humans miss fakes while AI catches them, but video compression makes detection harder. The detector can even identify which tampering method was used.
Dear fellow scholars, this is 2 Minute Papers with Zsolnai. We talked about the technique by the name Face2Face back in 2016, approximately 300 videos ago, it was able to take a video of us and transfer our gestures to a target subject. With techniques like this, it's now easier and cheaper than ever to create these deepfake videos of a target subject. Provided that we have enough training data. Where? Which is almost certainly the case for people who are the most high value targets for these kinds of operations. Look here. Some of these videos are real and some are fake. What do you think? Which is which? Well, here are the results. This one contains artifacts and is hence easy to spot. But the rest, it's tough and it's getting tougher by the day. How many did you get? Right, make sure to leave a comment below. However, don't despair, it's not all doom and gloom. Approximately a year ago in came Face Forensics, a paper that contains a large dataset of original and manipulated video pairs. As this offered a ton of training data for real and forged videos, it became possible to train a deepfake detector. You can see it here in action as these green to red colors showcase regions that the AI correctly thinks were tampered with. However, this follow up paper by the name Face Forensics contains not only an improved dataset, but provides many more valuable insights to help us detect these deepfake videos and even more. Let's dive in Key insight number one. As you've seen a minute ago, many of these deepfakes introduce imperfections artifacts to the video. However, most videos that we watch on the Internet are compressed and. And the compression procedure you have guessed right also introduces artifacts to the video. From this it follows that hiding these deepfake artifacts behind compressed videos sounds like a good strategy to fool humans and detector neural networks. Likewise. And not only that, but the paper also shows us by how much exactly. Here you see a table where each row shows the detection accuracy of previous techniques and a new proposed one. And the most interesting part is how this accuracy drops when we go from HQ to lq. Or in other words, from a high quality video to a lower quality one with more compression artifacts. Overall, we can get an 80 to 95% success rate, which is absolutely amazing. But of course you ask, amazing compared to what? Onwards to insight number two. This chart shows how humans fared in deepfake detection. And as you can see, not too well. Don't forget the 50% line means that the human guesses were as good as a coin flip, which means that they were not doing well at all. Face to face hovers around this ratio, and if you look at neural textures, you see that this is a technique that is extremely effective at fooling humans. And wait, what's that for? All the other techniques we see that the gray bars are shorter, meaning that it's more difficult to find out if a video is a deepfake because its own artifacts are hidden behind the compression artifacts. But the opposite is the case for neural textures, perhaps because of its small footprint on the videos. Note that the state of the art detector AI, for instance, the one proposed in this paper, does way better than these 204 human participants. This work does not only introduce a data set these cool insights, but also introduces a detector neural network. Now hold on to your papers, because this detection pipeline is not only so powerful that it practically eats compressed deepfakes for breakfast, but it even tells us with remarkable accuracy, which method was used to tamper with the input footage. Bravo. Now it is of utmost importance that we let the people know about the existence of these techniques. This is what I am trying to accomplish with this video, but that's not enough. So I also went to this year's biggest NATO conference and made sure that political and military decision makers are also informed about this topic. Last year I went to the European Political Strategy center with a similar goal. I was so nervous before both of these talks and spent a long time rehearsing them, which delayed a few videos here on the channel. However, because of your support on Patreon, I am in a fortunate situation where I can focus on doing what is right and what is the best for all of us and not worry about the financials all the time. I am really grateful for that. It really is a true privilege. Thank you. If you wish to support us, make sure to click the Patreon link in the video description. Thanks for watching and for your generous support and I'll see you next time.
