WEBVTT - TechStuff Classic: The National Facial Recognition Database 0:00:04.440 --> 0:00:10.400 Welcome to tech Stuff, a production from iHeartRadio. 0:00:12.400 --> 0:00:14.680 Hey there, and welcome to tech Stuff. I'm your host, 0:00:14.760 --> 0:00:15.520 John that Strickland. 0:00:15.520 --> 0:00:18.720 I'm an executive producer with iHeart Podcasts and how the 0:00:18.800 --> 0:00:22.400 tech are you? It's time for another classics episode. This 0:00:22.440 --> 0:00:27.080 episode originally published on May nineteenth, twenty seventeen. It is 0:00:27.160 --> 0:00:34.600 called the National Facial Recognition Database. Pretty, I would say, 0:00:34.680 --> 0:00:40.280 controversial topic. Well, let's listen in now. Before I dive 0:00:40.320 --> 0:00:41.919 into the topic, I want to make a couple of 0:00:41.920 --> 0:00:46.760 things very clear at the very beginning. First is I'm biased. 0:00:47.880 --> 0:00:50.720 I think the use of facial recognition software is problematic 0:00:50.880 --> 0:00:55.800 even if you have regulations in place. But I'm mostly 0:00:55.840 --> 0:01:00.480 talking about unregulated use because really we haven't a establish 0:01:00.600 --> 0:01:03.760 the rules and policies to guide the use of facial 0:01:03.800 --> 0:01:08.199 recognition software in a law enforcement context. So that's problem 0:01:08.280 --> 0:01:10.680 Number one is I have a very strong opinion about 0:01:10.680 --> 0:01:13.120 this and I'm not going to shy away from that. 0:01:15.560 --> 0:01:20.600 It's really unjustifiable to have unregulated use of facial recognition 0:01:20.920 --> 0:01:24.679 software in law enforcement contexts. So I want to make 0:01:24.720 --> 0:01:26.720 that clear out of the gate that I have this bias, 0:01:27.360 --> 0:01:30.360 and if that's an issue, that's fair, But at least 0:01:30.400 --> 0:01:32.880 I'm being honest, right, I'm not presenting this as if 0:01:32.920 --> 0:01:39.319 it's completely objective, unbiased information. I own this. You don't 0:01:39.360 --> 0:01:43.280 have to tell me. I know it already. Next, this 0:01:43.400 --> 0:01:47.680 is largely going to be a US centric discussion so 0:01:47.760 --> 0:01:51.280 that I can talk about details. But please know that 0:01:51.320 --> 0:01:53.600 there are a lot of these types of systems all 0:01:53.640 --> 0:01:56.600 over the world, not just in the United States, and 0:01:56.680 --> 0:01:59.040 a lot of these places have similar issues to the 0:01:59.080 --> 0:02:01.280 ones I'm going to be talking about here in the US. 0:02:01.960 --> 0:02:06.320 I'll just be focusing more on US stories to make 0:02:06.360 --> 0:02:09.720 specific points because this is where I live, and now 0:02:09.760 --> 0:02:12.560 to explain what I'm actually talking about here. So, back 0:02:12.600 --> 0:02:16.440 in twenty ten, the FBI undertook a project that cost 0:02:16.639 --> 0:02:20.399 more than an estimated one point two billion dollars that's 0:02:20.440 --> 0:02:24.160 billion with a B to replace what was called the 0:02:24.280 --> 0:02:29.560 Integrated Automated Fingerprint System or IAFS. Now, if I had 0:02:29.560 --> 0:02:32.720 been in place since nineteen ninety nine, and I've talked 0:02:32.760 --> 0:02:40.200 about fingerprints in a previous episode, IAFS was an attempt 0:02:40.400 --> 0:02:47.240 to create a usye database of fingerprint records so that 0:02:47.600 --> 0:02:50.000 if you were investigating a crime and you had lifted 0:02:50.040 --> 0:02:55.480 some prints from the crime, you could end up consulting 0:02:56.120 --> 0:02:58.600 this database and see if there are any matches in 0:02:58.639 --> 0:03:02.960 place to give you any leads on your investigation. The 0:03:03.160 --> 0:03:06.519 twenty ten project the FBI undertook was meant to vastly 0:03:06.680 --> 0:03:11.000 expand that capability by adding a lot more data to 0:03:11.240 --> 0:03:14.799 the database, not just fingerprints, but other stuff as well, 0:03:15.520 --> 0:03:19.360 and the new system is called the Next Generation Identification 0:03:19.639 --> 0:03:24.880 or NGI. It includes not just fingerprints, but other biographical 0:03:25.080 --> 0:03:30.200 data and biometrics information, including face recognition technology. So a 0:03:30.200 --> 0:03:34.680 lot of images are included in this particular database. So 0:03:34.720 --> 0:03:37.880 as part of this project, the FBI incorporated the Interstate 0:03:38.120 --> 0:03:42.960 Photo System or IPS, so you have NGI IPS it 0:03:43.080 --> 0:03:47.040 typically is how it's referred to now. That system includes 0:03:47.080 --> 0:03:51.080 images from police cases as well as photos from civil 0:03:51.240 --> 0:03:56.440 civic sources that are not necessarily related to crimes. That's 0:03:56.480 --> 0:03:59.480 not the only way the FBI can scan for a 0:03:59.560 --> 0:04:03.800 match of a photograph they've taken that relates to a 0:04:03.800 --> 0:04:07.119 case in some way to this massive database, but more 0:04:07.120 --> 0:04:10.880 on that in a little bit now. The general process 0:04:10.920 --> 0:04:14.800 of searching for a match follows a pretty simple pattern, 0:04:14.840 --> 0:04:18.479 although the details can be vastly different depending upon what 0:04:18.680 --> 0:04:23.200 facial recognition software you are using at the time. So 0:04:23.480 --> 0:04:26.680 you first start with an image related to a case, 0:04:27.160 --> 0:04:30.520 and this is called the probe photo. It is the 0:04:30.560 --> 0:04:34.280 one you are probing for lack of a better term, 0:04:35.920 --> 0:04:39.919 you don't know the identity of the person in the photograph, typically, 0:04:40.120 --> 0:04:42.400 or at least you might have suspicions, but you don't 0:04:42.440 --> 0:04:44.839 necessarily know for sure. So you've got a picture of 0:04:44.880 --> 0:04:50.000 an unknown person in this photograph. You then scan that 0:04:50.040 --> 0:04:53.800 photo and you use facial recognition software to analyze the 0:04:53.839 --> 0:04:56.800 picture and to try and find a match in this 0:04:57.000 --> 0:04:59.880 larger database. It starts searching all of the images in 0:04:59.880 --> 0:05:02.920 the database looking for any that might be a potential match. 0:05:03.560 --> 0:05:06.440 Depending upon the system and the policies that are in use, 0:05:06.880 --> 0:05:10.080 you could end up with a single photo return to you. 0:05:10.080 --> 0:05:13.080 You could end up with dozens of photos, so these 0:05:13.120 --> 0:05:16.520 would all be potential matches with different degrees of certainty 0:05:16.640 --> 0:05:19.680 for a match. You might remember in episodes I've talked 0:05:19.680 --> 0:05:22.640 about things like IBM's Watson that would come up with 0:05:22.800 --> 0:05:26.360 answers to a question and assign a value to each 0:05:26.400 --> 0:05:29.719 potential answer, and the one that had the highest value, 0:05:30.680 --> 0:05:34.640 assuming it's above a certain threshold, would be submitted as 0:05:34.720 --> 0:05:37.000 the answer. So it's not so much that the computer 0:05:37.080 --> 0:05:40.600 quote unquote knows it has a match. It suspects a 0:05:40.640 --> 0:05:43.960 match based upon a certain percentage as long as it's 0:05:44.040 --> 0:05:48.320 over a threshold of certainty, or you might end up 0:05:48.360 --> 0:05:50.960 with no photos at all. If no match was found 0:05:51.080 --> 0:05:55.640 or nothing ended up being above that threshold, the system 0:05:55.720 --> 0:05:58.440 might say, I couldn't match this photo with anyone who's 0:05:58.480 --> 0:06:04.080 in the database. A study performed by researchers at Georgetown 0:06:04.160 --> 0:06:09.840 University found that one in every two American adults has 0:06:09.880 --> 0:06:14.560 their face captured in an image database that is accessible 0:06:14.600 --> 0:06:19.200 by various law enforcement agencies, including but not limited to 0:06:19.360 --> 0:06:23.080 the IPS. In fact, the IPS has a small number 0:06:23.200 --> 0:06:27.800 of photos compared to the overall number represented by databases 0:06:27.839 --> 0:06:33.520 across the US. Now, this involves agencies at all different levels, federal, state, 0:06:33.640 --> 0:06:40.800 even tribal law for Native American tribes. That ends up 0:06:40.800 --> 0:06:45.520 being about one hundred and seventeen million people in these databases, 0:06:46.240 --> 0:06:50.160 many of whom, in fact large percentage of whom have 0:06:50.279 --> 0:06:54.200 no criminal background whatsoever. Their images are also in these databases, 0:06:54.600 --> 0:06:58.679 and this raises some big concerns about privacy and also accountability. 0:06:59.000 --> 0:07:01.880 So in today's episode, we're going to explore how facial 0:07:02.040 --> 0:07:07.480 recognition software works, as well as talk about the implementation 0:07:08.120 --> 0:07:12.000 for law enforcement and the reaction to this technology, and 0:07:12.040 --> 0:07:14.840 will probably listen to me get upset and a little 0:07:14.840 --> 0:07:18.400 head up about the whole thing in general. All right, 0:07:18.760 --> 0:07:22.200 So first, before we leap into the mess of law enforcement, 0:07:22.400 --> 0:07:26.640 because it is a mess, that's just a fact, let's 0:07:26.680 --> 0:07:30.880 talk first about the technology itself. When did facial recognition 0:07:31.000 --> 0:07:34.720 software get started and how does it work? Well, it's 0:07:34.760 --> 0:07:38.600 related to computer vision, which is a subset of artificial 0:07:38.640 --> 0:07:42.280 intelligence research. If you look at artificial intelligence, a lot 0:07:42.320 --> 0:07:45.240 of people simplify that by meaning, oh, this is so 0:07:45.320 --> 0:07:47.840 that you can teach computers how to think like people. 0:07:48.320 --> 0:07:51.520 But that's actually a very specific definition of a very 0:07:51.560 --> 0:07:54.840 specific type of artificial intelligence. When you really look at 0:07:54.840 --> 0:07:57.960 AI and you break it out, it involves a lot 0:07:57.960 --> 0:08:01.440 of subsets of abilities. One of those is the ability 0:08:01.440 --> 0:08:07.240 for machines to analyze imagery and be able to determine 0:08:07.280 --> 0:08:10.600 what that imagery represents. In a way, you could argue 0:08:10.640 --> 0:08:16.680 it's teaching computers how to understand pictures. It's also really challenging, 0:08:17.280 --> 0:08:20.360 and this is one of the object lessons that I 0:08:20.520 --> 0:08:25.320 use to teach people how Artificial intelligence is really tricky. 0:08:25.360 --> 0:08:28.920 It requires more than just pure processing power. I mean, 0:08:28.960 --> 0:08:32.720 processing power is important, but you can't solve all of 0:08:32.800 --> 0:08:36.600 AI's problems just by throwing more processors at it. You 0:08:36.679 --> 0:08:39.720 have to figure out from a software level how to 0:08:39.880 --> 0:08:43.599 leverage that processing power in a way that gives computers 0:08:43.640 --> 0:08:48.800 this ability to identify stuff based upon imagery. So a 0:08:48.840 --> 0:08:52.199 computer might be able to perform far more mathematical operations 0:08:52.240 --> 0:08:55.640 per second than even the cleverest of humans, but without 0:08:55.679 --> 0:08:58.040 the right software, they can't identify the picture of a 0:08:58.080 --> 0:09:01.960 seagull compared to say, a semi truck. You have to 0:09:02.040 --> 0:09:05.600 teach the computer how to do this. So let's say 0:09:05.640 --> 0:09:08.360 you develop a program that can analyze an image and 0:09:08.440 --> 0:09:13.720 break it down into simple data to describe that image, 0:09:13.960 --> 0:09:17.520 and then you essentially teach a computer what a coffee 0:09:17.559 --> 0:09:19.959 mug looks like. You take a picture of a coffee mug, 0:09:20.600 --> 0:09:23.960 you feed it to a computer, and you essentially say 0:09:24.280 --> 0:09:30.960 this data represents a coffee mug. You then would have 0:09:31.120 --> 0:09:36.200 to try and train the computer on what that actually means. 0:09:36.440 --> 0:09:39.640 The computer does not now know what a coffee mug is. 0:09:40.600 --> 0:09:44.560 It will recognize that specific mug in that specific orientation 0:09:44.840 --> 0:09:48.640 under those specific lighting conditions, assuming that you've designed the 0:09:48.640 --> 0:09:54.000 algorithm properly. But it's way more tricky than that. What 0:09:54.080 --> 0:09:56.680 if in the image that you fed the computer, the 0:09:56.720 --> 0:10:00.840 coffee mugs handle was facing to the left with respect 0:10:00.840 --> 0:10:04.040 of the viewer, but in a future picture the handle 0:10:04.160 --> 0:10:06.120 is off to the right instead of to the left, 0:10:06.200 --> 0:10:08.880 or it's turned around so you can't see the handle 0:10:08.920 --> 0:10:11.480 at all. It's behind the coffee mug. Well, if the 0:10:11.559 --> 0:10:14.600 mug is bigger or smaller, or a different shape, well 0:10:14.600 --> 0:10:18.200 if it's a different color. Image recognition is tough because 0:10:18.240 --> 0:10:23.280 computers don't immediately associate different objects within the same category 0:10:23.960 --> 0:10:28.400 as being the same thing. So if you teach me, Jonathan, 0:10:28.920 --> 0:10:31.240 what a coffee mug is, and you show me a 0:10:31.280 --> 0:10:34.840 couple of different examples saying, this is a coffee mug, 0:10:34.880 --> 0:10:37.080 but this is also a coffee mug, even though it's 0:10:37.080 --> 0:10:39.640 a different size and different shape and a different color, 0:10:40.320 --> 0:10:42.600 I'll catch on pretty quickly and it won't take very 0:10:42.600 --> 0:10:46.000 many coffee mugs for me to figure out. All Right, 0:10:46.040 --> 0:10:48.920 I got the basic idea of what a coffee mug is. 0:10:49.040 --> 0:10:52.280 I know what the concept of coffee mug is now, 0:10:53.000 --> 0:10:56.800 But computers aren't like that. You have to feed them 0:10:57.000 --> 0:11:00.600 thousands of images, both of coffee mugs and of not 0:11:01.000 --> 0:11:04.199 coffee mugs, so that the computer starts to be able 0:11:04.240 --> 0:11:08.920 to pick out the various features that are the essence 0:11:09.120 --> 0:11:12.520 of a coffee mug versus things that are not related 0:11:12.600 --> 0:11:16.400 to being a coffee mug. It takes hours and hours 0:11:16.400 --> 0:11:18.960 and hours of work of training these computers to do it, 0:11:19.000 --> 0:11:22.520 so it's a non trivial task, and this is true 0:11:22.559 --> 0:11:28.600 of all types of image recognition, including facial recognition. Now, 0:11:28.600 --> 0:11:34.480 to get around that problem, you end up sending thousands, 0:11:34.559 --> 0:11:38.080 countless thousands, millions maybe of images of what you're interested 0:11:38.120 --> 0:11:40.480 in while you're training the computer. And the nice thing 0:11:40.559 --> 0:11:43.920 is computers can process this information very very quickly, so 0:11:44.040 --> 0:11:49.520 while it takes a lot, it doesn't take relatively that long, 0:11:50.040 --> 0:11:52.959 it's not as laborious a process as it could be 0:11:53.120 --> 0:11:58.319 if computers were slower at analyzing information. So you might 0:11:58.360 --> 0:12:01.560 remember a story that kind of illustrates the point. Back 0:12:01.600 --> 0:12:05.520 in twenty twelve, there was a network of sixteen thousand 0:12:05.520 --> 0:12:11.080 computers that analyzed ten million images, and as a result, 0:12:11.080 --> 0:12:13.880 it could do the most important task any computer connected 0:12:13.880 --> 0:12:16.840 to the Internet should be expected to do. It could 0:12:16.840 --> 0:12:20.440 then identify cat videos because it now knew what a 0:12:20.440 --> 0:12:24.200 cat was, or at least the features that define catness. 0:12:25.080 --> 0:12:27.600 Catness as in the essence of being a cat, not 0:12:27.720 --> 0:12:31.520 a character from Hunger Games. Even then, there were times 0:12:31.520 --> 0:12:33.600 when a computer would get it wrong. Either it would 0:12:33.640 --> 0:12:35.880 not identify a cat as being a cat, or it 0:12:35.920 --> 0:12:38.880 would misidentify something else as being a cat because its 0:12:38.920 --> 0:12:41.360 features were close enough to cat like for it to 0:12:41.520 --> 0:12:46.360 fool the computer algorithm. A major breakthrough in facial recognition 0:12:46.400 --> 0:12:48.800 algorithms happened way back in two thousand and one. That's 0:12:48.840 --> 0:12:52.360 when Paul Viola and Michael Jones unveiled an algorithm for 0:12:52.440 --> 0:12:56.080 face detection, and it worked in real time, which meant 0:12:56.160 --> 0:12:59.600 that it could recognize a face that it would appear 0:12:59.679 --> 0:13:02.839 on a webcam. And by recognized, I mean it recognized 0:13:02.880 --> 0:13:06.840 that it was a face. It didn't assign an identity 0:13:07.360 --> 0:13:11.480 to the face. It didn't say, Oh, that's Bob, It said, oh, 0:13:11.679 --> 0:13:13.600 that is a face that is in front of the 0:13:13.600 --> 0:13:19.000 webcam right now. The algorithm soon found its way into 0:13:19.120 --> 0:13:24.040 open CV, which is an open source computer vision framework, 0:13:24.559 --> 0:13:28.079 and the open source approach allowed other programmers to dive 0:13:28.120 --> 0:13:31.000 into that code and to make changes and improvements, and 0:13:31.040 --> 0:13:36.360 it helped a rapid prototyping of facial recognition software to 0:13:36.520 --> 0:13:40.000 Other computer scientists who helped advance computer vision further were 0:13:40.080 --> 0:13:44.160 Bill Triggs and Navnit de Lal, who published a paper 0:13:44.160 --> 0:13:48.439 in two thousand and five about the histograbs of oriented gradients. Now, 0:13:48.480 --> 0:13:51.360 that was an approach that looked at gradient orientation in 0:13:51.400 --> 0:13:53.800 parts of an image, and essentially it describes the process 0:13:53.840 --> 0:13:56.960 of viewing an image with attention to edge directions and 0:13:57.000 --> 0:14:01.200 intensity gradients. That's a complicated way of saying the technique 0:14:01.240 --> 0:14:04.320 looks at the totality of a person, and then a 0:14:04.400 --> 0:14:07.640 machine learning algorithm determines whether or not that is actually 0:14:07.679 --> 0:14:11.600 a person or not a person. A bit later, computer 0:14:11.640 --> 0:14:15.520 scientists began pairing computer vision algorithms with deep learning and 0:14:15.679 --> 0:14:21.800 convolutional neural networks or CNNs. To go into this would 0:14:21.840 --> 0:14:25.040 require an episode all by itself. Neural networks are fascinating, 0:14:25.080 --> 0:14:28.160 but they're also pretty complicated, and I've got a whole 0:14:28.240 --> 0:14:31.400 lot of topics to cover today, so we can't really 0:14:31.440 --> 0:14:34.160 dive into it. You can think of an artificial neural 0:14:34.200 --> 0:14:38.600 network as designing a computer system that processes information in 0:14:38.640 --> 0:14:41.240 a way that's similar to the way our brains do. 0:14:41.720 --> 0:14:44.440 The computers are not thinking, but they are able to 0:14:44.480 --> 0:14:50.040 process information in a way that mimics how we process information, 0:14:50.520 --> 0:14:55.040 or a semi close approximation thereof that's a really kind 0:14:55.080 --> 0:14:57.000 of weak way of describing it. But again, to really 0:14:57.000 --> 0:15:03.360 go into detail will require a full episode all by itself. Typically, 0:15:03.640 --> 0:15:08.200 facial recognition software uses feature extraction to look for patterns 0:15:08.200 --> 0:15:11.680 in an image relating to facial features. In other words, 0:15:11.680 --> 0:15:15.440 it searches for features that resemble a face, the elements 0:15:15.440 --> 0:15:19.280 you would expect to be present in a typical face, 0:15:19.680 --> 0:15:24.280 So eyes, nose, a mouth, that would be major ones. Right. 0:15:24.560 --> 0:15:27.800 Then the software starts to estimate the relationships between those 0:15:27.840 --> 0:15:32.120 different elements. How wide are the eyes, how far apart 0:15:32.160 --> 0:15:34.040 are they from each other, How wide is the nose, 0:15:34.680 --> 0:15:37.640 how long is the jawline, what shape are the cheekbones? 0:15:39.000 --> 0:15:43.000 These sort of elements all play a part as points 0:15:43.000 --> 0:15:48.240 of data, and different facial recognition software packages weight these 0:15:48.280 --> 0:15:52.600 features in a different way. So it's not like I 0:15:52.640 --> 0:15:56.040 could say all facial recognition software looks at these four 0:15:56.080 --> 0:15:59.880 points of data as its primary source. It varies depending 0:15:59.920 --> 0:16:02.680 on upon the algorithm that's been designed by various companies, 0:16:03.480 --> 0:16:04.880 and part of the problem that we're going to talk 0:16:04.880 --> 0:16:09.200 about is that law enforcement across the United States they 0:16:09.200 --> 0:16:13.360 are not relying on a single facial recognition software approach. 0:16:13.640 --> 0:16:17.400 Different agencies have different vendors that they work with, So 0:16:18.640 --> 0:16:22.080 just because one might work very well doesn't necessarily mean 0:16:22.120 --> 0:16:25.400 it's competitors work just as well. And that's part of 0:16:25.400 --> 0:16:28.400 the problem. Now, all of these little points of data 0:16:28.400 --> 0:16:32.120 I'm talking about, these notle points and how they relate 0:16:32.160 --> 0:16:35.640 to one another, all of that gets boiled down into 0:16:35.720 --> 0:16:39.160 a numeric code that you could think of as a 0:16:39.200 --> 0:16:42.280 face print. This is supposed to be a representation of 0:16:42.320 --> 0:16:47.800 the unique set of data that is a compilation of 0:16:47.880 --> 0:16:53.480 all of these different points boiled down into numeric information itself. 0:16:56.080 --> 0:16:57.400 Then what you would do is you would have a 0:16:57.480 --> 0:17:02.520 database of face So if you wanted to find a match, 0:17:02.920 --> 0:17:05.920 you would feed the image you have, the probe image 0:17:06.080 --> 0:17:09.919 into this database, and the facial recognition software would analyze 0:17:09.960 --> 0:17:13.240 the probe photo. It would end up assigning this numeric 0:17:13.400 --> 0:17:16.480 value and would start looking through the database for other 0:17:16.560 --> 0:17:20.040 numeric values that were as similar to that probe one 0:17:20.160 --> 0:17:25.800 as possible and start returning those images as potential matches 0:17:26.119 --> 0:17:29.560 or candidates. They tend to use the word candidate photos. 0:17:30.640 --> 0:17:33.199 Otherwise you'll either get no match at all or you 0:17:33.240 --> 0:17:35.800 get a false positive. You will end up getting an 0:17:35.800 --> 0:17:39.919 image of someone who looks like the person whose image 0:17:39.920 --> 0:17:44.480 you submitted, but is not the same person. That does happen, 0:17:45.240 --> 0:17:48.200 And that's the basic way that facial recognition software works. 0:17:49.119 --> 0:17:51.439 But keep in mind, different vendors use all their own 0:17:51.480 --> 0:17:54.560 specific approaches, like I said, and some could be less 0:17:54.600 --> 0:17:58.840 accurate than others. Some might be accurate for specific ethnicities 0:17:58.880 --> 0:18:01.600 and not as accurate as other ones. That's a huge problem, 0:18:03.240 --> 0:18:08.000 so it gets complicated. Even when I'm talking in more 0:18:08.040 --> 0:18:10.960 general terms, you have to remember that there are a 0:18:10.960 --> 0:18:18.359 lot of specific incidents and specific implementations of facial recognition 0:18:18.480 --> 0:18:23.080 software that have their own issues. So I'm gonna be 0:18:23.160 --> 0:18:24.920 as general as I can. I'm not going to call 0:18:24.960 --> 0:18:29.280 out any particular facial recognition software vendors out there. I'm 0:18:29.280 --> 0:18:32.879 more going to talk about the overall issues that various 0:18:32.960 --> 0:18:38.040 organizations have had as they've looked into this topic. Now, 0:18:38.080 --> 0:18:40.879 there are plenty of applications for facial recognition that have 0:18:40.920 --> 0:18:43.119 nothing to do with identifying a person. I mentioned that 0:18:43.200 --> 0:18:45.840 earlier that there was the one for a webcam that 0:18:45.880 --> 0:18:48.320 could identify when a face was in front of the webcam. 0:18:48.440 --> 0:18:51.720 This wasn't to identify anybody. It was again just to say, yes, 0:18:51.760 --> 0:18:55.080 there's somebody looking into the webcam at this moment, which 0:18:55.119 --> 0:18:57.560 by itself can be useful and have nothing to do 0:18:57.600 --> 0:19:01.400 with identification. There are plenty of digital cameras out there 0:19:01.480 --> 0:19:06.199 and camera phone apps that can identify when there's a 0:19:06.280 --> 0:19:10.000 face looking at the camera, and again it's not necessarily 0:19:10.000 --> 0:19:12.440 to identify that person, but rather to say, oh, well, 0:19:12.720 --> 0:19:15.880 this is a face. The camera is most likely trying 0:19:15.880 --> 0:19:18.520 to focus on this person, so let's make this person 0:19:18.560 --> 0:19:21.560 the point of focus and not focus on something in 0:19:21.600 --> 0:19:25.119 the background like a tree that's fifty yards back. Instead, 0:19:25.160 --> 0:19:28.240 let's focus on the person who's in the foreground. So 0:19:28.280 --> 0:19:33.560 that's pretty handy, and again there's nothing particularly problematic from 0:19:33.600 --> 0:19:36.320 an identification standpoint, because that's not the purpose of it. 0:19:38.119 --> 0:19:41.919 But then you also have other implementations, like on social media, 0:19:42.200 --> 0:19:45.240 which allow you to do things like tag people based 0:19:45.320 --> 0:19:50.080 upon an algorithm recognizing a person. So Facebook is a 0:19:50.080 --> 0:19:52.560 great example of this. Right, if you upload a picture 0:19:52.560 --> 0:19:55.960 of one of your Facebook friends onto Facebook chances are 0:19:56.000 --> 0:19:59.080 it's giving you a suggestion to tag that photo with 0:19:59.280 --> 0:20:04.600 the specific in mind. That may not be that problematic either, 0:20:05.200 --> 0:20:08.640 depending upon how your friend feels about pictures being uploaded 0:20:08.640 --> 0:20:13.959 to Facebook. Some people are very cautious about that, and 0:20:14.240 --> 0:20:16.399 of course you know, I always recommend you talk to 0:20:16.440 --> 0:20:20.200 anybody before you start tagging folks on Facebook photos, just 0:20:20.240 --> 0:20:22.679 to make sure they're fine with it. I say that 0:20:22.720 --> 0:20:25.320 as a person who has done it, and then notice 0:20:25.359 --> 0:20:27.440 that some of my tags got removed by the people 0:20:27.480 --> 0:20:30.920 I tagged later on, which taught me I should probably 0:20:31.000 --> 0:20:35.760 ask first, rather than give them the feeling that they 0:20:35.760 --> 0:20:39.080 need to go and remove a tag or two. We've 0:20:39.080 --> 0:20:43.120 also seen examples of this simple implementation of facial recognition 0:20:43.200 --> 0:20:49.080 going awry. Google's street View will blur out faces, for example, 0:20:49.560 --> 0:20:53.520 in an effort to protect people's identity while street view 0:20:53.520 --> 0:20:56.399 cars are out and about taking images. This makes sense. 0:20:56.680 --> 0:20:58.920 Let's say that you are in a part of town 0:20:59.280 --> 0:21:01.840 that you normally would not be in. For whatever reason, 0:21:02.040 --> 0:21:05.080 you might not want your picture to be included on 0:21:05.119 --> 0:21:07.800 Google street View, so that whenever anyone looks at that 0:21:07.880 --> 0:21:11.200 street for that point forward, they see your face on there, 0:21:11.800 --> 0:21:15.639 you know, coming out of I don't know a Wendy's. 0:21:15.880 --> 0:21:19.320 Maybe you are a manager for burger King that would 0:21:19.320 --> 0:21:23.359 look bad, or you know, lots of other reasons that 0:21:23.520 --> 0:21:26.600 obviously can spring to mind as well. You don't want 0:21:26.640 --> 0:21:32.200 to violate someone's privacy. But Google StreetView would also blur 0:21:32.320 --> 0:21:35.400 out images that were not real people faces, like images 0:21:35.440 --> 0:21:38.440 on billboards or murals. Sometimes if it had a person's 0:21:38.480 --> 0:21:40.920 face on a mural, the face would be blurred out, 0:21:40.920 --> 0:21:42.960 even though it's not a real person, it's just a 0:21:43.000 --> 0:21:47.320 painting or In September twenty sixteen, c Neet reported on 0:21:47.359 --> 0:21:49.760 an incident in which Google street View blurred out the 0:21:49.760 --> 0:21:53.840 face of a cow. So Google was being very thoughtful 0:21:53.960 --> 0:22:00.439 to protect that cow's privacy. But what about matching faces 0:22:00.440 --> 0:22:04.679 to identities? So in some cases, again seemingly harmless if 0:22:04.720 --> 0:22:07.280 you want to tag your friends, but when it comes 0:22:07.320 --> 0:22:10.600 to law enforcement, things get a bit sticky, particularly as 0:22:10.640 --> 0:22:13.160 you learn more about the specifics. And we'll talk about 0:22:13.160 --> 0:22:16.240 that in just a second, but first let's take a 0:22:16.320 --> 0:22:28.119 quick break to thank our sponsor. All right, let's first 0:22:28.160 --> 0:22:33.720 start with the FBI's Interstate Photos System, or IPS, because 0:22:33.800 --> 0:22:37.399 this one has perhaps the least controversial elements to it 0:22:37.440 --> 0:22:40.679 when you really look at it, it's still problematic, but 0:22:40.880 --> 0:22:45.560 not nearly as much as the larger picture. The system 0:22:45.600 --> 0:22:51.880 contains both images from criminal cases like mugshots and things 0:22:51.960 --> 0:22:54.920 of that nature, but it also includes some photos from 0:22:55.080 --> 0:23:00.880 civil sources like ID applications, that kind of thing. When 0:23:00.920 --> 0:23:04.760 the Government Accountability Office or GAO, they're gonna be a 0:23:04.760 --> 0:23:08.720 lot of acronyms and initializations or initialisms, I should say 0:23:08.800 --> 0:23:11.520 in this episode, so I apologize for that. But Government 0:23:11.600 --> 0:23:16.480 Accountability Office they did a study on this matter just 0:23:16.920 --> 0:23:20.680 in twenty sixteen, so not that long ago. They published 0:23:20.680 --> 0:23:24.480 its report on facial recognition software use among law enforcements, 0:23:24.520 --> 0:23:28.919 specifically the FBI because they're a federal agency, so they 0:23:28.960 --> 0:23:33.719 were concerned with the federal use of this. The database 0:23:34.040 --> 0:23:37.240 contained about thirty million photos at the time of the 0:23:37.359 --> 0:23:41.680 GAO study, so thirty million pictures are in this database. 0:23:42.119 --> 0:23:46.440 Most of those images came from eighteen thousand different law 0:23:46.520 --> 0:23:50.720 enforcement agencies at all levels of government, that includes the 0:23:50.720 --> 0:23:55.840 tribal law enforcement offices. About seventy percent of all the 0:23:55.840 --> 0:24:01.119 photos in the database were mugshots. More than of the 0:24:01.119 --> 0:24:06.080 photos in that database are from criminal cases, so that 0:24:06.119 --> 0:24:10.040 means that less than twenty percent were from civil sources. 0:24:10.920 --> 0:24:15.120 In addition to that, there were some cases, plenty of them, 0:24:15.480 --> 0:24:19.760 where the database had images of people both from a 0:24:19.760 --> 0:24:23.520 civil source and from a criminal source. So I'll give 0:24:23.560 --> 0:24:27.320 you a theoretical example. Let's say that sometime in the 0:24:27.320 --> 0:24:33.840 past I got nabbed by the cops for grand theft 0:24:33.840 --> 0:24:37.840 auto because I play that game. But let's say that 0:24:37.880 --> 0:24:40.080 I stole a car, which we already know is a 0:24:40.160 --> 0:24:43.920 complete fabrication because I don't even drive. But let's say 0:24:43.920 --> 0:24:47.199 I stole a car, and that I had moved the 0:24:47.200 --> 0:24:51.440 car across state lines. It became a federal case. Therefore, 0:24:51.960 --> 0:24:55.920 my criminal information is included. My mugshot would be included 0:24:56.160 --> 0:25:03.879 in this particular database. On related note, my ID also 0:25:04.720 --> 0:25:08.639 is in that database as a civil image, not as 0:25:08.640 --> 0:25:11.800 a criminal image. Well, in my case, they would tie 0:25:11.840 --> 0:25:15.920 those two images together because they refer to the same 0:25:16.000 --> 0:25:19.200 person and I had been involved in a criminal act. 0:25:20.119 --> 0:25:23.080 So while I would have an image in there from 0:25:23.119 --> 0:25:26.440 a civil source, it would be filed under the criminal 0:25:26.480 --> 0:25:28.720 side of things. This is important when we get to 0:25:28.960 --> 0:25:32.920 how the probes work. Now, let's say you have been 0:25:33.400 --> 0:25:37.840 perfectly law abiding this whole time, and that your ID 0:25:39.000 --> 0:25:41.840 is also in this database, but it's just under the 0:25:41.840 --> 0:25:45.200 civil side of things. Since you don't have any criminal background, 0:25:45.840 --> 0:25:49.600 it's not connected to anything on the criminal side, So 0:25:49.760 --> 0:25:54.160 when it comes to probes using the IPS, your information 0:25:54.680 --> 0:25:59.520 will not be referenced because the FBI policy is when 0:25:59.520 --> 0:26:03.720 it's running these potential matches with a photo that's been 0:26:03.800 --> 0:26:06.800 gathered as part of the evidence for an ongoing investigation, 0:26:07.520 --> 0:26:11.639 they can only consult the criminal side, not the civil side, 0:26:12.280 --> 0:26:15.960 with the exception of any civil photos that are connected 0:26:16.000 --> 0:26:20.280