WEBVTT - The National Facial Recognition Database 0:00:04.200 --> 0:00:07.240 Get in text with technology with tech Stuff from how 0:00:07.280 --> 0:00:14.080 stuff works dot com. Hey there, and welcome to text Stuff. 0:00:14.120 --> 0:00:18.520 I'm your host Jonathan Strickland, senior writer with how stuff 0:00:18.600 --> 0:00:22.400 works dot com, and today we're going to examine a 0:00:22.520 --> 0:00:27.240 controversial topic, namely how law enforcement is using facial recognition 0:00:27.280 --> 0:00:31.240 software and the problems that this raises. Now, before I 0:00:31.320 --> 0:00:33.040 dive into the topic, I want to make a couple 0:00:33.080 --> 0:00:37.239 of things very clear at the very beginning. First, is 0:00:37.320 --> 0:00:41.159 I'm biased. I think the use of facial recognition software 0:00:41.240 --> 0:00:46.400 is problematic even if you have regulations in place. But 0:00:46.479 --> 0:00:50.960 I'm mostly talking about unregulated use because really we haven't 0:00:51.200 --> 0:00:54.680 established the rules and policies to guide the use of 0:00:54.720 --> 0:00:58.640 facial recognition software in a law enforcement context. So that's 0:00:59.120 --> 0:01:01.640 problem No. One is. I have a very strong opinion 0:01:01.640 --> 0:01:06.800 about this, and I'm not gonna shy away from that. Um, 0:01:06.840 --> 0:01:11.880 it's really unjustifiable to have unregulated use of facial recognition 0:01:12.160 --> 0:01:15.920 software in law enforcement contexts. So I want to make 0:01:15.959 --> 0:01:18.000 that clear out of the gate, that I have this bias, 0:01:18.600 --> 0:01:21.600 and if that's an issue, that's fair. But at least 0:01:21.640 --> 0:01:24.160 I'm being honest, right, I'm not presenting this as if 0:01:24.160 --> 0:01:30.560 it's completely objective, unbiased information. I own this. You don't 0:01:30.600 --> 0:01:34.600 have to tell me. I know it already. Next, this 0:01:34.680 --> 0:01:39.000 is largely going to be a US centric discussion so 0:01:39.040 --> 0:01:42.520 that I can talk about details. But please know that 0:01:42.600 --> 0:01:44.880 there are a lot of these types of systems all 0:01:44.880 --> 0:01:47.840 over the world, not just in the United States, and 0:01:47.920 --> 0:01:50.360 a lot of these places have similar issues to the 0:01:50.360 --> 0:01:52.480 ones I'm gonna be talking about here in the US. 0:01:53.200 --> 0:01:57.560 I'll just be focusing more on US stories to make 0:01:57.600 --> 0:02:00.960 specific points because this is where I live, and now 0:02:01.000 --> 0:02:03.800 to explain what I'm actually talking about here. So back 0:02:03.840 --> 0:02:08.280 in the FBI undertook a project that cost more than 0:02:08.320 --> 0:02:12.200 an estimated one point two billion dollars that's billion with 0:02:12.280 --> 0:02:16.840 a B to replace what was called the Integrated Automated 0:02:17.040 --> 0:02:20.799 Fingerprint System or I A f S. Have I only 0:02:20.800 --> 0:02:24.919 been in place since nine and I've talked about fingerprints 0:02:25.600 --> 0:02:29.960 in a previous episode. UM. The i F I a 0:02:30.200 --> 0:02:34.880 f S was an attempt to create a A U 0:02:35.080 --> 0:02:39.160 S Y database of fingerprint records so that if you 0:02:39.200 --> 0:02:41.799 were investigating a crime and you had lifted some prints 0:02:41.800 --> 0:02:47.480 from the crime, you could end up UH consulting this 0:02:47.639 --> 0:02:50.160 database and see if there are any matches in place 0:02:50.200 --> 0:02:55.520 to give you any leads on your investigation. The project 0:02:55.560 --> 0:02:59.200 the FBI undertook was meant to vastly expand that capability 0:02:59.240 --> 0:03:03.560 by adding a lot more data to the database, not 0:03:03.639 --> 0:03:07.239 just fingerprints, but other stuff as well, and the news 0:03:07.320 --> 0:03:11.840 system is called the Next Generation Identification or in g I. 0:03:12.960 --> 0:03:16.920 It includes not just fingerprints, but other biographical data and 0:03:17.040 --> 0:03:21.720 biometrics information, including face recognition technology. So a lot of 0:03:21.840 --> 0:03:26.280 images are included in this particular database. So as part 0:03:26.280 --> 0:03:30.400 of this project, the FBI incorporated the Interstate Photo System 0:03:30.560 --> 0:03:33.000 or i p S, so you have n g I 0:03:33.120 --> 0:03:37.360 dash ips that typically is how it's referred to now. 0:03:37.360 --> 0:03:40.720 That system includes images from police cases as well as 0:03:40.920 --> 0:03:45.640 photos from civil civic sources that are not necessarily related 0:03:45.680 --> 0:03:49.520 to crimes. Not that's not the only way the FBI 0:03:49.640 --> 0:03:53.680 can scan for a match of a photograph they've taken 0:03:53.880 --> 0:03:56.720 that relates to a case in some way to this 0:03:56.840 --> 0:04:01.080 massive database, But more on that in a little bit now. 0:04:01.080 --> 0:04:04.880 The general process of searching for a match follows a 0:04:04.880 --> 0:04:08.800 pretty simple pattern, although the details can be vastly different 0:04:08.880 --> 0:04:12.360 depending upon what facial recognition software you are using at 0:04:12.360 --> 0:04:17.040 the time. So you first start with an image related 0:04:17.240 --> 0:04:20.799 to a case, and this is called the probe photo. 0:04:21.200 --> 0:04:24.880 It is the one you are probing for lack of 0:04:24.880 --> 0:04:28.640 a better term, You don't know the identity of the 0:04:28.640 --> 0:04:32.320 person in the photograph, typically, or at least you might 0:04:32.360 --> 0:04:35.279 have suspicions, but you don't necessarily know for sure. So 0:04:35.320 --> 0:04:40.039 you've got a picture of an unknown person in this photograph. 0:04:40.360 --> 0:04:43.560 You then scan that photo and you use facial recognition 0:04:43.560 --> 0:04:46.880 software to analyze the picture and to try and find 0:04:46.920 --> 0:04:50.480 a match in this larger database. It starts searching all 0:04:50.520 --> 0:04:52.760 of the images in this database looking for any that 0:04:52.839 --> 0:04:56.279 might be a potential match. Depending upon the system and 0:04:56.320 --> 0:04:59.080 the policies that are in use, you could end up 0:04:59.080 --> 0:05:01.640 with a single photo do returned to you. You could 0:05:01.760 --> 0:05:04.760 end up with dozens of photos. So these would all 0:05:04.760 --> 0:05:08.479 be potential matches with different degrees of certainty for a match. 0:05:08.920 --> 0:05:11.560 You might remember in episodes I've talked about things like 0:05:11.600 --> 0:05:14.599 IBM S Watson that would come up with answers to 0:05:14.640 --> 0:05:18.520 a question and assign a value to each potential answer, 0:05:19.080 --> 0:05:22.640 and the one that had the highest value, assuming it's 0:05:22.640 --> 0:05:26.560 above a certain threshold, would be submitted as the answer. 0:05:26.960 --> 0:05:29.000 So it's not so much that the computer quote unquote 0:05:29.120 --> 0:05:32.520 knows it has a match. It suspects a match based 0:05:32.560 --> 0:05:36.159 upon a certain percentage as long as it's over a 0:05:36.279 --> 0:05:39.839 threshold of certainty, or you might end up with no 0:05:39.960 --> 0:05:42.839 photos at all. If no match was found or nothing 0:05:43.320 --> 0:05:47.400 ended up being above that threshold, the system might say, 0:05:47.480 --> 0:05:50.520 I couldn't match this photo with anyone who's in the database. 0:05:52.680 --> 0:05:56.720 A study performed by researchers at Georgetown University found that 0:05:57.279 --> 0:06:02.520 one in every two American adults has their face captured 0:06:03.080 --> 0:06:06.839 in an image database that is accessible by various law 0:06:06.920 --> 0:06:11.400 enforcement agencies, including but not limited to, the i p S. 0:06:11.440 --> 0:06:14.320 In fact, the i p S has a small number 0:06:14.440 --> 0:06:19.040 of photos compared to the overall number represented by databases 0:06:19.080 --> 0:06:24.800 across the US. Now, this involves agencies at all different levels, federal, state, 0:06:24.880 --> 0:06:31.880 even tribal law for Native American uh tribes. That ends 0:06:31.920 --> 0:06:36.760 up being about a hundred seventeen million people in these databases, 0:06:37.480 --> 0:06:41.359 many of whom, in fact large personage of whom have 0:06:41.520 --> 0:06:45.440 no criminal background whatsoever. Their images are also in these databases, 0:06:45.839 --> 0:06:49.920 and this raises some big concerns about privacy and also accountability. 0:06:50.240 --> 0:06:53.160 So in today's episode, we're going to explore how facial 0:06:53.279 --> 0:06:57.960 recognition software works. As well as talk about the implement 0:06:58.040 --> 0:07:03.120 implementation for law and forcement and the reaction to this technology, 0:07:03.200 --> 0:07:05.760 and will probably listen to me get upset and a 0:07:05.800 --> 0:07:09.600 little head up about the whole thing in general. All Right, 0:07:10.000 --> 0:07:13.440 So first, before we leap into the mess of law enforcement, 0:07:13.640 --> 0:07:17.880 because it is a mess, that's just a fact, let's 0:07:17.920 --> 0:07:22.200 talk first about the technology itself. When did facial recognition 0:07:22.240 --> 0:07:26.000 software get started and how does it work? Well, it's 0:07:26.000 --> 0:07:29.880 related to computer vision, which is a subset of artificial 0:07:29.920 --> 0:07:33.520 intelligence research. If you look at artificial intelligence, a lot 0:07:33.560 --> 0:07:36.560 of people simplify that by meaning, oh, this is so 0:07:36.600 --> 0:07:39.080 that you can teach computers how to think like people. 0:07:39.560 --> 0:07:42.760 But that's actually a very specific definition of a very 0:07:42.800 --> 0:07:46.080 specific type of artificial intelligence. When you really look at 0:07:46.120 --> 0:07:49.040 a I and you break it out, and involves a 0:07:49.080 --> 0:07:52.080 lot of subsets of abilities, and one of those is 0:07:52.120 --> 0:07:57.800 the ability for machines to analyze imagery and be able 0:07:57.840 --> 0:08:01.200 to determine what that imagery represent. In a way, you 0:08:01.240 --> 0:08:06.480 could argue it's teaching computers how to understand pictures. It's 0:08:06.520 --> 0:08:10.320 also really challenging, and this is one of the object 0:08:10.400 --> 0:08:14.960 lessons that I use to teach people how artificial intelligence 0:08:15.680 --> 0:08:19.400 is really tricky. It requires more than just pure processing power. 0:08:19.920 --> 0:08:23.520 I mean, processing power is important, but you can't solve 0:08:23.680 --> 0:08:26.920 all of AI's problems just by throwing more processors at it. 0:08:27.760 --> 0:08:30.760 You have to figure out from a software level how 0:08:30.840 --> 0:08:34.240 to leverage that processing power in a way that gives 0:08:34.280 --> 0:08:39.840 computers this ability to identify stuff based upon imagery. So 0:08:40.000 --> 0:08:42.800 a computer might be able to perform far more mathematical 0:08:42.840 --> 0:08:46.640 operations per second than even the cleverest of humans, but 0:08:46.679 --> 0:08:49.160 without the right software, they can't identify the picture of 0:08:49.200 --> 0:08:53.160 a seagull compared to say, a semitruck. You have to 0:08:53.280 --> 0:08:56.880 teach the computer how to do this. So let's say 0:08:56.880 --> 0:08:59.600 you develop a program that can analyze an image and 0:08:59.640 --> 0:09:05.520 breaking down into simple data to describe that image, and 0:09:05.600 --> 0:09:09.080 then you essentially teach a computer what a coffee mug 0:09:09.160 --> 0:09:11.280 looks like. You take a picture of a coffee mug, 0:09:11.880 --> 0:09:15.479 you feed it to a computer, and you essentially say 0:09:15.520 --> 0:09:22.240 this data represents a coffee mug. You then would have 0:09:22.360 --> 0:09:27.439 to try and train the computer on what that actually means. 0:09:27.679 --> 0:09:30.840 The computer does not now know what a coffee mug is. 0:09:31.880 --> 0:09:35.920 It will recognize that specific mug in that specific orientation 0:09:36.080 --> 0:09:39.880 under those specific lighting conditions, assuming that you've designed the 0:09:39.920 --> 0:09:45.240 algorithm properly. But it's way more tricky than that. What 0:09:45.360 --> 0:09:47.920 if in the image that you fed the computer, the 0:09:48.000 --> 0:09:52.080 coffee mugs handle was facing to the left with respect 0:09:52.120 --> 0:09:55.320 of the viewer, but in a future picture the handle 0:09:55.400 --> 0:09:57.360 is off to the right instead of to the left, 0:09:57.480 --> 0:10:00.000 or it's turned around so you can't see the hand 0:10:00.040 --> 0:10:02.720 to at all. It's behind the coffee mug. What if 0:10:02.720 --> 0:10:05.480 the mug is bigger or smaller, or a different shape, 0:10:05.720 --> 0:10:09.160 what if it's a different color. Image recognition is tough 0:10:09.200 --> 0:10:13.880 because computers don't immediately associate different objects within the same 0:10:13.920 --> 0:10:19.640 category as being the same thing. So if you teach me, Jonathan, 0:10:20.200 --> 0:10:22.520 what a coffee mug is, and you show me a 0:10:22.520 --> 0:10:26.080 couple of different examples saying, this is a coffee mug, 0:10:26.120 --> 0:10:28.360 but this is also a coffee mug, even though it's 0:10:28.360 --> 0:10:30.839 a different size, and different shape and a different color, 0:10:31.600 --> 0:10:33.839 I'll catch on pretty quickly and it won't take very 0:10:33.880 --> 0:10:37.200 many coffee mugs for me to figure out. All Right, 0:10:37.280 --> 0:10:40.120 I got the basic idea of what a coffee mug is. 0:10:40.280 --> 0:10:43.480 I know what the concept of coffee mug is now, 0:10:44.280 --> 0:10:48.040 but computers aren't like that. You have to feed them 0:10:48.280 --> 0:10:51.880 thousands of images, both of coffee mugs and of not 0:10:52.240 --> 0:10:55.440 coffee mugs, so that the computer starts to be able 0:10:55.480 --> 0:10:59.960 to pick out the various features that are the say 0:11:00.040 --> 0:11:03.280 it's of a coffee mug versus things that are not 0:11:03.360 --> 0:11:07.320 related to being a coffee mug. It takes hours and 0:11:07.360 --> 0:11:09.920 hours and hours of work of training these computers to 0:11:09.960 --> 0:11:13.079 do it, so it's a non trivial task, and this 0:11:13.160 --> 0:11:19.839 is true of all types of image recognition, including facial recognition. Now, 0:11:19.880 --> 0:11:25.720 to get around that problem, you end up sending thousands, 0:11:25.800 --> 0:11:29.319 countless thousands, millions maybe of images of what you're interested 0:11:29.360 --> 0:11:31.719 in while you're training the computer. And the nice thing 0:11:31.800 --> 0:11:35.240 is computers can process this information very very quickly, so 0:11:35.280 --> 0:11:40.800 while it takes a lot, it doesn't take relatively that long. 0:11:41.280 --> 0:11:44.199 It's not as laborious a process as it could be 0:11:44.360 --> 0:11:49.559 if computers were slower at analyzing information. So you might 0:11:49.600 --> 0:11:52.800 remember a story that kind of illustrates this point. Back 0:11:52.840 --> 0:11:56.240 in two thousand twelve, there was a network of sixteen 0:11:56.320 --> 0:12:01.680 thousand computers that analyzed ten million only in images, and 0:12:01.760 --> 0:12:03.880 as a result, it could do the most important task 0:12:04.000 --> 0:12:07.160 any computer connected to the Internet should be expected to do. 0:12:07.760 --> 0:12:11.400 It could then identify cat videos because it now knew 0:12:11.400 --> 0:12:14.240 what a cat was, or at least the features that 0:12:14.320 --> 0:12:18.199 defined cat nous catness as in the essence of being 0:12:18.200 --> 0:12:22.280 a cat, not a character from Hunger Games. Even then, 0:12:22.320 --> 0:12:24.000 there were times when the computer would get it wrong. 0:12:24.200 --> 0:12:26.840 Either it would not identify a cat as being a cat, 0:12:27.040 --> 0:12:30.000 or it misidentify something else as being a cat because 0:12:30.040 --> 0:12:32.440 its features were close enough to cat like for it 0:12:32.480 --> 0:12:37.120 to fool the computer algorithm. A major breakthrough and facial 0:12:37.160 --> 0:12:40.080 recognition algorithms happened way back in two thousand one. That's 0:12:40.080 --> 0:12:43.640 when Paul Viola and Michael Jones unveiled an algorithm for 0:12:43.679 --> 0:12:47.360 face detection, and it worked in real time, which meant 0:12:47.400 --> 0:12:50.880 that it could recognize a face that it would appear 0:12:51.000 --> 0:12:53.440 on a webcam. And but I recognize I mean it 0:12:53.520 --> 0:12:57.439 recognized that it was a face. It didn't assign an 0:12:57.480 --> 0:13:01.160 identity to the face. It didn't say, oh, that's Bob, 0:13:01.600 --> 0:13:04.400 It said, oh, that is a face that is in 0:13:04.440 --> 0:13:09.000 front of the webcam right now. Uh. The algorithm soon 0:13:09.000 --> 0:13:12.520 found its way into open CV, which is an open 0:13:12.600 --> 0:13:17.160 source computer vision framework, and the open source approach allowed 0:13:17.200 --> 0:13:20.920 other programmers to dive into that code and to make 0:13:21.040 --> 0:13:24.679 changes and improvements, and it helped a rapid prototyping of 0:13:24.679 --> 0:13:29.480 facial recognition software to other computer scientists who helped advance 0:13:29.520 --> 0:13:33.680 computer vision further. Where Bill Triggs and Novnique de Lal 0:13:34.520 --> 0:13:36.560 who published the paper in two thousand five about the 0:13:36.640 --> 0:13:40.800 histograms of oriented gradients. Now, that was an approach that 0:13:40.840 --> 0:13:43.480 looked at gradient orientation in parts of an image, and 0:13:43.600 --> 0:13:46.520 essentially it describes the process of viewing an image with 0:13:46.559 --> 0:13:51.200 attention to edge directions and intensity gradients. That's a complicated 0:13:51.240 --> 0:13:54.000 way of saying the technique looks at the totalitary t 0:13:54.320 --> 0:13:57.240 of a person, and then a machine learning algorithm determines 0:13:57.280 --> 0:13:59.600 whether or not that is actually a person or not 0:13:59.640 --> 0:14:04.079 a person, and a bent later, computer scientists began pairing 0:14:04.080 --> 0:14:08.880 computer vision algorithms with deep learning and convolutional neural networks 0:14:08.960 --> 0:14:13.880 or CNNs. To go into this would require an episode 0:14:13.880 --> 0:14:16.880 all by itself. Neural networks are fascinating, but they're also 0:14:17.080 --> 0:14:20.560 pretty complicated, and I've got a whole lot of topics 0:14:20.560 --> 0:14:23.400 to cover today, so we can't really dive into it. 0:14:23.920 --> 0:14:27.520 You can think of an artificial neural network as designing 0:14:27.520 --> 0:14:30.480 a computer system that processes information in a way that's 0:14:30.680 --> 0:14:33.720 similar to the way our brains do. The computers are 0:14:33.720 --> 0:14:37.600 not thinking, but they are able to process information in 0:14:37.640 --> 0:14:41.960 a way that mimics how we process information, or a 0:14:42.040 --> 0:14:46.320 close a semi close approximation thereof that's a really kind 0:14:46.320 --> 0:14:48.240 of weak way of describing it. But again, to really 0:14:48.240 --> 0:14:54.600 go into detail will require a full episode of by itself. Typically, 0:14:54.880 --> 0:14:59.440 facial recognition software uses feature extraction to look for patterns 0:14:59.440 --> 0:15:02.920 in an image relating to facial features. In other words, 0:15:02.920 --> 0:15:06.080 it's searches for for features that resemble a face, the 0:15:06.240 --> 0:15:10.520 elements you would expect to be present in a typical face, 0:15:10.960 --> 0:15:15.520 so eyes, knows, a mouth, that would be major ones. Right. 0:15:15.800 --> 0:15:19.040 Then the software starts to estimate the relationships between those 0:15:19.080 --> 0:15:23.360 different elements. How wide are the eyes, how far apart 0:15:23.400 --> 0:15:25.320 are they from each other, How white is the nose, 0:15:25.920 --> 0:15:28.920 how long is the jawline, what shape are the cheap bones? 0:15:30.240 --> 0:15:34.680 These sort of elements all play apart as points of data, 0:15:36.000 --> 0:15:40.520 and different facial recognition software packages weight these features in 0:15:40.560 --> 0:15:44.080 a different way. So it's it's not like I could 0:15:44.080 --> 0:15:47.640 say all facial recognition software looks at these four points 0:15:47.640 --> 0:15:51.480 of data as its primary source. It varies depending upon 0:15:51.520 --> 0:15:54.840 the algorithm that's been designed by various companies. UH And 0:15:54.880 --> 0:15:56.560 part of the problem that we're going to talk about 0:15:56.680 --> 0:16:00.800 is that law enforcement across the United States, they're not 0:16:00.880 --> 0:16:05.920 relying on a single facial recognition software approach. Different agencies 0:16:06.000 --> 0:16:10.400 have different vendors that they work with, So just because 0:16:10.480 --> 0:16:14.280 one might work very well doesn't necessarily mean it's competitors 0:16:14.320 --> 0:16:17.920 work just as well. And that's part of the problem. Now, 0:16:18.040 --> 0:16:20.680 all of these little points of data I'm talking about, 0:16:20.720 --> 0:16:24.080 these nodal points and how they relate to one another, 0:16:24.200 --> 0:16:28.160 all of that gets boiled down into a numeric code 0:16:28.840 --> 0:16:31.600 that you could think of as a face print. This 0:16:31.720 --> 0:16:34.520 is supposed to be a representation of the unique set 0:16:34.520 --> 0:16:39.640 of data that is a compilation of all of these 0:16:39.640 --> 0:16:47.680 different points boiled down into numeric information itself. Then what 0:16:47.760 --> 0:16:51.520 you would do is you would have a database of faces. 0:16:52.560 --> 0:16:54.520 So if you want to find a match, you would 0:16:54.560 --> 0:16:58.479 feed the image you have, the probe image into this database, 0:16:58.920 --> 0:17:02.080 and the facial recognition software would analyze the probe photo. 0:17:02.760 --> 0:17:05.680 It would end up assigning this numeric value and would 0:17:05.720 --> 0:17:08.960 start looking through the database for other numeric values that 0:17:09.040 --> 0:17:12.800 were as similar to that probe one as possible, and 0:17:12.920 --> 0:17:19.119 start returning those images as potential matches or candidates. They 0:17:19.160 --> 0:17:23.200 tend to use the word candidate photos. Otherwise you'll either 0:17:23.240 --> 0:17:25.520 get no match at all or you get a false positive. 0:17:25.760 --> 0:17:28.000 You will end up getting an image of someone who 0:17:28.040 --> 0:17:32.600 looks like the person whose image you submitted, but is 0:17:32.720 --> 0:17:36.919 not the same person. That does happen, and that's the 0:17:36.960 --> 0:17:41.040 basic way that facial recognition software works. But keep in mind, 0:17:41.040 --> 0:17:44.240 different vendors use all their own specific approaches, like I said, 0:17:44.760 --> 0:17:47.480 and some could be less accurate than others. Some might 0:17:47.520 --> 0:17:51.159 be accurate for specific ethnicities and not as accurate as 0:17:51.200 --> 0:17:57.400 other ones. That's a huge problem. So it gets complicated. 0:17:57.960 --> 0:18:01.040 Even when I'm talking in more general terms, you have 0:18:01.080 --> 0:18:05.120 to remember that there are a lot of specific UH 0:18:05.359 --> 0:18:11.720 incidents and specific implementations of facial recognition software that have 0:18:11.880 --> 0:18:15.119 their own issues. So I'm gonna be as general as 0:18:15.160 --> 0:18:17.160 I can. I'm not going to call out any particular 0:18:17.200 --> 0:18:21.119 facial recognition software vendors out there. I'm more going to 0:18:21.160 --> 0:18:25.320 talk about the overall issues that various organizations have had 0:18:25.359 --> 0:18:29.960 as they've looked into this topic. Now, there are plenty 0:18:30.000 --> 0:18:32.600 of applications for facial recognition that have nothing to do 0:18:32.640 --> 0:18:35.640 with identifying a person. I mentioned that earlier that there 0:18:35.720 --> 0:18:38.040 was the one for a webcam that could identify when 0:18:38.040 --> 0:18:40.160 a face was in front of the webcam. This wasn't 0:18:40.200 --> 0:18:42.960 to identify anybody. It was again just to say, yes, 0:18:43.040 --> 0:18:46.359 there's somebody looking into the webcam at this moment, which 0:18:46.359 --> 0:18:48.800 by itself can be useful and have nothing to do 0:18:48.880 --> 0:18:52.680 with identification. There are plenty of digital cameras out there 0:18:52.760 --> 0:18:57.359 and and camera phone apps that can identify when there's 0:18:57.400 --> 0:19:01.240 a face looking at the camera. Again, it's not necessarily 0:19:01.280 --> 0:19:03.760 to identify that person, but rather to say, oh, well, 0:19:03.960 --> 0:19:07.120 this is a face. The camera is most likely trying 0:19:07.119 --> 0:19:09.760 to focus on this person, so let's make this person 0:19:09.840 --> 0:19:12.840 the point of focus and not focus on something in 0:19:12.840 --> 0:19:16.399 the background like a tree that's fifty yards back. Instead, 0:19:16.440 --> 0:19:19.520 let's focus on the person who's in the foreground. So 0:19:19.520 --> 0:19:24.800 that's pretty handy, and again there's nothing particularly problematic from 0:19:24.840 --> 0:19:27.560 an identification standpoint, because that's not the purpose of it. 0:19:29.400 --> 0:19:33.159 But then you also have other implementations, like on social media, 0:19:33.480 --> 0:19:36.520 which allow you to do things like tag people based 0:19:36.600 --> 0:19:41.080 upon a an algorithm recognizing a person. So Facebook is 0:19:41.240 --> 0:19:43.480 a great example of this. Right, if you upload a 0:19:43.520 --> 0:19:47.040 picture of one of your Facebook friends onto Facebook. Chances 0:19:47.040 --> 0:19:49.720 are it's giving you a suggestion to tag that photo 0:19:50.240 --> 0:19:54.040 with the specific person in mind. That may not be 0:19:54.520 --> 0:19:58.679 that problematic either, depending upon how your friend feels about 0:19:58.760 --> 0:20:03.159 pictures being uploaded to Facebook. Some people are very um 0:20:03.200 --> 0:20:06.800 cautious about that, And of course you know, I always 0:20:06.800 --> 0:20:09.760 recommend you talk to anybody before you start tagging folks 0:20:10.240 --> 0:20:13.440 on Facebook photos, just to make sure they're fine with it. Um. 0:20:13.480 --> 0:20:15.320 I say that as a person who has done it, 0:20:15.840 --> 0:20:18.120 and then notice that some of my tags got removed 0:20:18.160 --> 0:20:21.199 by the people I tagged later on, which taught me 0:20:21.400 --> 0:20:26.359 I should probably ask first, rather than give them the 0:20:26.400 --> 0:20:28.240 feeling that they need to go and remove a tag 0:20:28.359 --> 0:20:33.400 or two. We've also seen examples of this simple implementation 0:20:33.440 --> 0:20:38.800 of facial recognition going awry. Google's street View will blur 0:20:38.880 --> 0:20:42.360 out faces, for example, in an effort to protect people's 0:20:42.400 --> 0:20:46.520 identity while street view cars are out and about taking images. 0:20:46.920 --> 0:20:49.359 This makes sense. Let's let's say that you are in 0:20:49.400 --> 0:20:52.000 a part of town that you normally would not be in. 0:20:52.119 --> 0:20:55.080 For whatever reason, you might not want your picture to 0:20:55.200 --> 0:20:58.560 be included on Google street View, so that whenever anyone 0:20:58.640 --> 0:21:01.480 looks at that street for that point forward. They see 0:21:01.520 --> 0:21:05.080 your face on there, you know, coming out of uh, 0:21:05.200 --> 0:21:08.680 I don't know, a Windy's. Maybe you are a manager 0:21:08.760 --> 0:21:12.919 for Burger King that would look bad, or you know, 0:21:13.440 --> 0:21:16.520 lots of other reasons that obviously can spring to mind 0:21:16.560 --> 0:21:20.639 as well. You don't want to violate someone's privacy. But 0:21:21.600 --> 0:21:24.679 Google StreetView would also blur out images that were not 0:21:24.800 --> 0:21:28.720 real people faces, like images on billboards or murals. Sometimes 0:21:28.720 --> 0:21:31.119 if it had a person's face on a mural, the 0:21:31.160 --> 0:21:32.840 face would be blurred out, even though it's not a 0:21:32.840 --> 0:21:38.000 real person, it's just a painting. Or In September, Set 0:21:38.080 --> 0:21:40.800 reported on an incident in which Google street View blurred 0:21:40.800 --> 0:21:44.080 out the face of a cow. So Google was being 0:21:44.240 --> 0:21:49.879 very thoughtful to protect that cow's privacy. But what about 0:21:50.720 --> 0:21:55.160 matching faces to identities? So in some cases, again seemingly 0:21:55.200 --> 0:21:58.240 harmless if you want to tag your friends, but when 0:21:58.280 --> 0:22:00.600 it comes to law enforcement, things get a bit sticky, 0:22:01.080 --> 0:22:03.960 particularly as you learn more about the specifics. And we'll 0:22:04.000 --> 0:22:07.240 talk about that in just a second, but first let's 0:22:07.280 --> 0:22:17.679 take a quick break to thank our sponsor. All right, 0:22:17.760 --> 0:22:22.520 let's first start with the FBI's Interstate Photos System or 0:22:22.560 --> 0:22:26.000 i p S, because this one has perhaps the least 0:22:26.119 --> 0:22:29.800 controversial elements to it. When you really look at it, 0:22:29.800 --> 0:22:33.280 it's still problematic, but not nearly as much as the 0:22:33.440 --> 0:22:38.760 larger picture. The system contains both images from criminal cases 0:22:39.800 --> 0:22:43.280 like mug shots UH and things of that nature, but 0:22:43.359 --> 0:22:47.760 it also includes some photos from civil sources like UH, 0:22:48.040 --> 0:22:51.560 like I D applications, that kind of thing. When the 0:22:51.600 --> 0:22:55.159 Government Accountability Office or g a O, there's gonna be 0:22:55.200 --> 0:22:59.000 a lot of acronyms and initializations are initialisms, I should 0:22:59.040 --> 0:23:01.320 say in this episodes, so I apologize for that. But 0:23:01.720 --> 0:23:06.000 Government Accountability Office they did a study on this matter 0:23:06.840 --> 0:23:11.240 just in sten So not that long ago. They published 0:23:11.240 --> 0:23:15.040 its report on facial recognition software use among law enforcements, 0:23:15.080 --> 0:23:19.479 specifically the FBI because they're a federal agency, so they 0:23:19.520 --> 0:23:24.240 were concerned with the federal use of this. The database 0:23:24.560 --> 0:23:27.800 contained about thirty million photos at the time of the 0:23:27.880 --> 0:23:31.439 g a O study, so thirty million pictures are in 0:23:31.480 --> 0:23:35.960 this database. Most of those images came from eighteen thousand 0:23:36.240 --> 0:23:40.639 different law enforcement agencies at all levels of government, that 0:23:40.680 --> 0:23:46.240 includes the tribal law enforcement offices. About sevent of all 0:23:46.280 --> 0:23:50.120 the photos and the database were mug shots. UH. More 0:23:50.160 --> 0:23:53.879 than eighty percent of the photos in that database are 0:23:53.960 --> 0:23:58.040 from criminal cases, so that means that less than twenty 0:23:58.600 --> 0:24:03.880 were from civil sources. In addition to that, there were 0:24:03.920 --> 0:24:07.800 some cases, plenty of them, where the database had images 0:24:08.160 --> 0:24:11.760 of people both from a civil source and from a 0:24:11.760 --> 0:24:16.720 criminal source. So I'll give you a theoretical example. Let's 0:24:16.720 --> 0:24:20.840 say that sometime in the past, UH, I got nabbed 0:24:20.880 --> 0:24:26.000 by the cops for for grand theft auto because I 0:24:26.080 --> 0:24:29.080 play that game. But let's say that I stole a car, 0:24:29.240 --> 0:24:33.000 which we already know is a complete fabrication because I 0:24:33.040 --> 0:24:35.120 don't even drive. But let's say I stole a car 0:24:35.480 --> 0:24:39.080 and that I had moved the car across state lines. 0:24:39.640 --> 0:24:45.000 It became a federal case. Therefore, my criminal information is included. 0:24:45.040 --> 0:24:49.639 My mug shot would be included in this particular database. UH. 0:24:50.359 --> 0:24:55.600 On a related note, my I d also is in 0:24:55.640 --> 0:24:59.240 that database as a civil UH image, not as a 0:24:59.240 --> 0:25:02.600 criminal image. Well, in my case, they would tie those 0:25:02.600 --> 0:25:06.919 two images together because they refer to the same person 0:25:07.160 --> 0:25:10.880 and I had been involved in a criminal act. So 0:25:11.200 --> 0:25:13.720 while I would have an image in there from a 0:25:13.800 --> 0:25:17.280 civil source, it would be filed under the criminal side 0:25:17.280 --> 0:25:19.680 of things. This is important when we get to how 0:25:19.800 --> 0:25:24.600 the probes work. Now let's say you have been perfectly 0:25:25.000 --> 0:25:28.359 law abiding this whole time, and that your I D 0:25:29.600 --> 0:25:32.320 is also in this database, but it's just under the 0:25:32.440 --> 0:25:35.760 civil side of things. Since you don't have any criminal background, 0:25:36.400 --> 0:25:40.200 it's not connected to anything on the criminal side. So 0:25:40.280 --> 0:25:43.240 when it comes to probes using the i p S, 0:25:43.920 --> 0:25:49.120 your information will not be referenced. Because the FBI policy 0:25:49.720 --> 0:25:53.359 is when it's running these these potential matches with a 0:25:53.359 --> 0:25:56.000