Speaker 1: Okay. What percent likelihood do you think Chinese Elon Musk is a
Speaker 2: Deg. I would say that this is 90% big. My name is Ji and I'm a research director at the information sciences Institute. And I'm also the director of the visual intelligence and multimedia analytics laboratory at ISI. We have been working for the last few years on developing different generations for detecting deep defects. We create artificial intelligence algorithms [00:00:30] that take as an input, any given video, whether this video is fake or really, since we don't know the artificial intelligence algorithm produces a number between zero and one, indicating whether this video is actually fake or real, let's have a look at some videos and see whether we can actually judge whether they're fake or real
Speaker 3: A fellow Americans.
Speaker 2: So the first video is a video of former president, Donald Trump, [00:01:00] the audio and the lips are not imperfect sync play using any video player and slow it down. A little bit.
Speaker 3: Many of you will have been
Speaker 2: Followed at a lower speed. You can clearly see that the audio and lips are not very well synced. And the other thing that you can look at is where the president's neck meets the, the core of the shirt. It's clearly a phase super imposed on a previous video. And this is the output of running the same video through the second generation [00:01:30] of Al's, uh, defect detector. This indicator between R and fake indicates that this particular frame is actually fake. Meanwhile, the indicator on top of the video here indicates that most of the video is actually fake. The second video we'll talk about is where the creator actually super imposed can or Reeves picture on first gum. As you can see here, if you start noticing the facial expressions, especially on the forehead, do not change much from one frame to the other, indicating [00:02:00] that this is actually, uh, a fake video.
Speaker 2: And now let's look at the output. As you can see here with this bar next to the face, it shows that this is a deep fake video. The third video is actually an extremely challenging video. This is a Kindra Lamar music video. And as you can see here, now, the creator of the video replaced Kindra Lamar's face with OG Simpson, me as an expert in that domain. If I didn't know that this is actually a deep fake, [00:02:30] I wouldn't have known that this is a deep fake, because it's extremely well done. And now the creator of the video replaces even OJ, Samsung, again with Kanye. Now, let me play you. What's the output of the algorithm. This video was very well made using, uh, extremely advanced computer graphics and AI technology, and neither a human observer or even the algorithm were able to identify the defect parts of, of that video.
Speaker 2: So this is the part where the face has been replaced, OJ Simpson. [00:03:00] And as you can see here, the, the algorithm actually thinks it's a real video, cause it has been really well made. If I have to take a random shot, why the algorithm didn't actually work on this? It's because of a very well known challenge and AI community now, which is the bias of AI algorithms against underrepresented groups, including black people. Initially, when people started collecting data sets for face recogni mission or defect, et cetera, nobody was paying attention to the imbalance of the dataset. So [00:03:30] the imbalance of these dataset causes these BI biases in any AI algorithm, the unfairness and bias of AI algorithms is a very well known problem in the AI community now. And so many people, including our lab here are working to trying to solve and assess, uh, this problem.
Speaker 4: Our enemies can make it look like anyone is saying anything at any point in
Speaker 2: Time. The following video has been a really famous video for deep defects where only the lips and facial expressions of president Obama have been replaced in what used to [00:04:00] be an authentic video. As the video says, this is clearly a deep fake as admitted by the video. So Jordan peel actually created this video as an educational tool to let everybody understand the dangers of deep fake. So as you can see here, the algorithm actually successfully identified pretty much all of the video as, as deep fake. The next video is where, uh, the creator of the video replaces the face of Jim Carey with actor William DFU. The deep fake here is actually very [00:04:30] well made. And I personally don't have any idea how it was created. The algorithm detects parts of the defects and misses some parts of the defects. If you notice here, when the lighting is not that good or the face of the actor is not very well lit, the algorithm struggles to find whether this is a deep fake or a real video here, the face is slightly will lit and the algorithm can successfully identify this as a deep defect.
Speaker 5: So much of our wow today comes to us through these
Speaker 2: Screens. The next video is actually from the [00:05:00] United Kingdom where somebody created a video of the queen, a really easy way to figure out that this is a deep defect is to look at the, the chin of the queen, where the video is clearly a little bit blurred compared to the rest of the face, which is an indication that the, the person who created, uh, this video had difficulty blending the deep fake face with the rest of the body. So as you can see here, this is a professionally made video for educational purposes. [00:05:30] There is a green background it's, uh, it's taking in a studio, which makes it extremely difficult to, to detect for algorithms, but let's look at the output of the algorithm anyway. So surprisingly, actually the algorithm does really good job identifying that most of these frames of the queen are actually deep fit.
Speaker 2: So as the camera gets closer to the queen, the algorithm does slightly better job. We cannot expect the algorithm will work well on every single frame of the video, but overall, the algorithm does a pretty good job identifying [00:06:00] that most of the video is actually a deep fit. The next video is from TikTok from China. The video is actually extremely well made. I don't know what kind of technology was used to create this video. Before I show you the output of the algorithm. Let me show you how to identify that this is actually a deep fake, just as a human observer. And let me just slow the video a little bit, look at where he starts putting back his sunglasses on his face. You can see there is a little bit of a jump. The algorithm has a little bit of a difficulty at [00:06:30] a few frames that are not, not noticeable at high speed.
Speaker 2: And now let me play the output of the deep fake detection algorithm. So the aggressive actually does a really decent job figuring out that this is a deep fake giving that it's very well made. Big chunks of the video are identified as deep fakes. Whenever you see a brown part of the bar here, this essentially means this is a deep fake. The rest video is also from TikTok. The creator of the video did a really, really good job creating the video, trying to make [00:07:00] people think that there is actually a Chinese Elon Musk looklike overall. The algorithm did a really good job identifying that pretty much. Most of the video is, is deep fake, except for these very two very small parts of the video. Every single human being and every single algorithm out there will always make mistakes in terms of false alarms and detection breaks nobody or no algorithm will ever miss everything and catch everything.
Speaker 2: And this is an extremely important point since even if [00:07:30] we cannot catch every single defect out there, even if we can just catch 60% of them, that's still better than nothing detecting 50 or 60% of defects out there is definitely much better than letting the line between truth and fake. Just alone defects on their own are not harmful until they go viral on social networks and people start believing them. And once people start believing in defects, you cannot roll that back. So before you reshare think carefully about [00:08:00] whether this particular defects will be harmful to the society and play it more than once slowed down to make sure that it actually is a real video.
