WEBVTT 00:00.720 --> 00:04.000 In this video we're going to take a look at eye detection. 00:04.280 --> 00:08.440 Now I am deepfake videos. 00:08.480 --> 00:14.760 I manipulated videos and photos and even audio is a growing concern. 00:16.400 --> 00:23.840 There's been cases where very large banks have been tricked into sending over millions and millions 00:23.840 --> 00:34.280 of dollars, and this was done by AI video with AI voice over a video call convincing the bank to actually 00:34.280 --> 00:39.480 transfer funds because they thought they were talking to, because it looked like they were talking 00:39.520 --> 00:43.120 to someone from a one of their clients. 00:44.600 --> 00:46.600 And it was very convincing looking. 00:46.600 --> 00:47.120 Video. 00:47.240 --> 00:50.040 The audio matched up, but it was all AI generated. 00:50.800 --> 00:55.360 Now there's also photo manipulation using AI. 00:55.840 --> 01:04.330 It's so easy generating AI photos also, and we see this a lot with social media posts, posting pictures, 01:04.330 --> 01:12.210 and basically convincing people of things that aren't true, including manipulation for political gains, 01:12.410 --> 01:18.170 for causing discourse in society and whatnot. 01:19.130 --> 01:21.970 So how do we figure out if it's a R or not? 01:22.010 --> 01:27.170 Well, there are certain things that we could do in terms of video chats with AI. 01:28.610 --> 01:35.530 You could actually, um, and this will work up until AI gets to a point where it's actually has better 01:35.530 --> 01:36.250 tracking. 01:37.130 --> 01:43.330 There's been cases where people had interviews with other people, and the person was using AI to kind 01:43.370 --> 01:44.930 of hide who they really were. 01:45.610 --> 01:51.050 So things, uh, if you're doing video chatting, you could say, hey, okay, I need you to stand up 01:51.050 --> 01:51.410 real quick. 01:51.410 --> 01:53.370 I need you to turn your head left and right real fast. 01:53.370 --> 01:56.010 I need you to, you know, walk to the side. 01:56.210 --> 01:58.130 Uh uh, quickly walk to the left. 01:58.130 --> 01:59.210 Walk to the right real quick. 01:59.210 --> 02:07.030 And in in those cases, I generally at this point has problems keeping up with real quick movements, 02:07.430 --> 02:08.310 sudden movements. 02:08.310 --> 02:09.950 So you'll see that break in there. 02:10.670 --> 02:19.430 And in cases where people actually have interviewed someone that was using I and they, they use that, 02:19.990 --> 02:25.230 those people usually drop off of the interview because they don't want to get caught, or in some cases 02:25.230 --> 02:26.030 they actually do it. 02:26.030 --> 02:29.030 And you could see it, uh, the video break. 02:30.190 --> 02:39.150 Now, other ways we could do this is we could use things like, uh, this, this website at TV, Audio 02:40.390 --> 02:40.910 and 02:41.670 --> 02:45.870 video.tv.com. 02:46.990 --> 02:51.990 Um, I have the full one here because I'm logged in here and you can register for a free account. 02:51.990 --> 02:54.350 You could do five scans a month for free. 02:55.150 --> 02:58.950 Now, what this does is it'll analyze the video. 02:58.950 --> 03:02.080 So either we could drop a video in here or we could do a URL. 03:02.600 --> 03:05.520 So I'm going to drop a video in here and I'm using this one here. 03:05.520 --> 03:07.160 This is from YouTube. 03:07.160 --> 03:16.120 This is a basically someone re-imagined the Avengers using actors from the 80s. 03:16.960 --> 03:22.720 So we're going to drop it in here and we're going to see basically how this does. 03:22.760 --> 03:24.320 So I have that input here. 03:24.320 --> 03:25.320 Video is ready to scan. 03:25.360 --> 03:30.320 I'm going to click on Scan Video here and we'll let this run through. 03:34.520 --> 03:36.680 And this is actually going pretty quickly. 03:36.680 --> 03:39.800 So we'll get a result pretty fast here. 03:47.920 --> 03:48.520 Okay. 03:48.680 --> 03:50.640 So we'll let that scan through. 03:50.880 --> 03:59.010 And I do want to take the opportunity to say that as with every tool I recommend if you can do multiple 03:59.010 --> 04:03.810 scans with different tools to kind of verify your results. 04:03.810 --> 04:13.850 And just because we're using a tool to specifically look for eye videos, these things are not always 04:13.890 --> 04:14.330 accurate. 04:14.330 --> 04:16.530 It doesn't always do the best job detecting it. 04:16.530 --> 04:22.930 That's why I like using multiple tools to run through different scans, not just with eye, but with 04:22.970 --> 04:23.890 other stuff too. 04:23.930 --> 04:25.170 Basically anything. 04:25.770 --> 04:27.730 I like to verify the results. 04:27.730 --> 04:31.290 So this came back with a 61 score. 04:31.290 --> 04:32.770 So it says suspicious. 04:32.770 --> 04:35.970 And we could see the video analysis here. 04:36.890 --> 04:39.410 And you could actually go in here. 04:39.410 --> 04:48.970 You could see estimate score lip placement face placement generated 95%, confidence interval 50%. 04:49.490 --> 04:51.890 And you could try to read that chart. 04:52.210 --> 04:56.810 So I'm just going to scroll down here because I'm not really great reading that chart. 04:56.810 --> 05:01.870 So let's scroll down in here and it says technical suspicious rating. 05:01.870 --> 05:03.230 So it's a three out of four. 05:03.950 --> 05:04.630 Now it's a show. 05:04.670 --> 05:10.630 Technical suspicion is rating of 61 which means that parts of the video are suspicious. 05:10.870 --> 05:16.950 So again this is I'll actually play the video for make sure this is muted. 05:17.390 --> 05:18.990 So this is what it looks like. 05:20.110 --> 05:27.510 So the still pictures it actually looks pretty good uh for being AI generated. 05:27.550 --> 05:34.630 It it looks like it could very well be the, uh, the actual actors in most scenes, but you could see 05:34.670 --> 05:40.390 kind of some funny things about, about some of the transitioning, some of the movements because it's 05:40.390 --> 05:40.750 AI. 05:40.990 --> 05:47.670 But in terms of the face itself, they do a pretty good job with, um, how the faces were generated. 05:48.110 --> 05:55.670 But again, this is a tool that we could potentially use to check for AI generated content. 05:56.230 --> 05:57.390 So thank you for watching. 05:57.390 --> 05:58.630 I'll see you next video.