[00:07.310 --> 00:12.010]  Hi, my name is Richard Smith, and I run a computer site called Computer Bytes Man.
[00:12.770 --> 00:15.230]  This is the second talk for me.
[00:15.290 --> 00:18.150]  I did a talk yesterday on tracking criminals on the internet.
[00:19.190 --> 00:20.590]  Today is going to be a little bit different.
[00:20.730 --> 00:30.010]  It's more of a gadget talk, if you will, but it's going to get into some sort of practical applications of gadgets as well as some policy decisions that are going to be made.
[00:30.010 --> 00:36.570]  This is an area that I've been following for more than a year and a half.
[00:37.750 --> 00:59.950]  I originally got into looking at face recognition software when reading the news stories about what got called the Snooper Bowl, the Super Bowl that was not this year's but the previous year's in Tampa, Florida, had a face recognition set up so that each person who went through the turnstile was scanned by a face recognition system.
[01:00.010 --> 01:07.110]  And there was a database of about 3,000 people in the Tampa area who wanted various kinds of criminal warrants.
[01:07.510 --> 01:12.630]  And so they used the Super Bowl as a place to round up the local criminals.
[01:12.790 --> 01:15.190]  The cops found it as a convenient place to do this.
[01:16.070 --> 01:20.450]  Well, they claim they got 19 matches but no arrests.
[01:21.630 --> 01:29.890]  The company said it worked, but I think most people's calculations of it really didn't because no one was arrested.
[01:30.690 --> 01:36.990]  Anyway, the system was put in though with the idea of keeping us safe from the terrorists.
[01:37.170 --> 01:39.610]  This was before September 11th even.
[01:40.630 --> 01:44.610]  The company decided to issue a press release that this had happened.
[01:44.610 --> 01:49.390]  And then sort of created a bit of controversy, more than they probably expected.
[01:50.270 --> 01:56.370]  It actually got a Big Brother award from Privacy International which I actually gave the nomination for.
[01:59.350 --> 02:11.570]  After that system, the next big controversy with face scanning was again in the Tampa area where the local police again wanted to round up the usual suspects.
[02:11.570 --> 02:18.330]  And put a face scanning system in Ybor City which is sort of the entertainment area of Tampa.
[02:18.590 --> 02:21.530]  So the nightclub and restaurant area, sort of Spanish area.
[02:23.750 --> 02:31.930]  And the same idea was to equip the regular closed circuit TV systems which are throughout that neighborhood.
[02:32.990 --> 02:36.670]  Equip them with face recognition so that they could find the local criminals.
[02:37.870 --> 02:40.330]  And again, it's done as a convenience for the police.
[02:40.370 --> 02:41.870]  It just makes them easier to spot people.
[02:42.170 --> 02:47.870]  And they have a... in that particular system they have a license to put up 30,000 people in the database.
[02:47.930 --> 02:50.030]  Although they only have a few thousand so far.
[02:50.590 --> 02:58.090]  I actually spoke to the... the head... the guy at the Tampa Police Department was head of this project.
[02:58.390 --> 03:00.630]  And he gave me the rundown of who was in there.
[03:00.830 --> 03:01.630]  And it was kind of interesting.
[03:01.830 --> 03:06.350]  It was, you know, people who wanted on outstanding criminal warrants.
[03:07.050 --> 03:08.430]  Known sex offenders.
[03:08.990 --> 03:17.350]  And just general bad people that have been arrested before for, you know, pickpocketing or whatever that they thought might hang out there.
[03:18.510 --> 03:23.670]  One of the interesting things is they could arrest the people who are wanting on outstanding arrest warrants.
[03:24.050 --> 03:29.710]  But if someone was in the sexual offender registry and wasn't supposed to be in the neighborhood, they weren't allowed to arrest them.
[03:30.390 --> 03:32.750]  All they could do is come up to them and say, you better move along.
[03:33.270 --> 03:35.310]  Because they didn't have jurisdiction over these people.
[03:35.470 --> 03:40.250]  So there's sort of... these systems have sort of all sorts of interesting policy decisions with them.
[03:41.550 --> 03:49.070]  However, face skating or face recognition software became a really hot item after September 11th.
[03:49.630 --> 03:52.890]  And it's an interesting study in commerce of why that happened.
[03:53.810 --> 03:59.590]  There's basically, here in the United States, two companies who are the big players in the face recognition software business.
[03:59.910 --> 04:03.450]  One is Visage, and they're up in the Boston area around where I live.
[04:03.450 --> 04:07.850]  And the other company is Physionics, which is now known as Identics.
[04:08.330 --> 04:13.610]  And they have offices actually over in Jersey City as well as a couple other places around the country.
[04:14.090 --> 04:25.670]  And basically, what happened is that both the companies saw a business opportunity after 9-11 to offer security services for airports.
[04:25.670 --> 04:34.210]  And so they came up with this idea of using face recognition systems, equipping every airport with them, to look for the bad terrorists.
[04:35.770 --> 04:40.670]  And I became kind of skeptic on this on a lot of different levels.
[04:40.710 --> 04:48.830]  And what I hope to do today is sort of impart some of that skepticism here of what I see as a problem with these systems.
[04:49.690 --> 05:00.770]  And it's very interesting, you know, they're very sort of high-techy products in the sense that they take video images and attempt to match up features on people's faces with databases.
[05:00.990 --> 05:02.930]  So there's a lot of sort of interesting technology into them.
[05:03.670 --> 05:13.530]  But there's also a lot of very interesting real-world problems that one must deal with when installing these systems and consider, you know, how they can really be used.
[05:13.530 --> 05:24.810]  And it's an interesting study also in how, in times of sort of troubles here with terrorism, that, you know, companies are sort of taking advantage of the situation and making pitches on these products.
[05:32.480 --> 05:36.080]  First of all, what are some of the uses for the face-scanning systems at airports?
[05:36.220 --> 05:45.340]  One is the one that's getting all the press and the one that's making, like, you know, Peter Jennings's evening news and so on is spotting suspected terrorists in the terminals themselves.
[05:45.900 --> 05:57.080]  So the basic idea is you have, you take your CCTVs or other video cameras you've set up, and you have a database of known terrorists, and you're looking for them at the airport.
[05:57.740 --> 06:08.820]  And the idea is you try to spot them at the airport, and if you see somebody in the database, it's visually confirmed by a person, and if, by darn, if it's really one of those terrorists, then you go tackle them in the terminal.
[06:08.820 --> 06:09.780]  That's the basic idea.
[06:11.780 --> 06:15.620]  Another thing that can be done that's being talked about is in some background checks.
[06:15.780 --> 06:26.620]  So before you get on the airplane, they figure out, they, again, the same thing, spotting the terrorists, but they're the ones who are actually trying to get on the airplanes.
[06:28.540 --> 06:34.380]  A more traditional application for face scanning would be access control to restricted areas.
[06:34.580 --> 06:45.280]  You know, today at the airport, you know, we're all familiar with the swipe cards that the employees use at the airport, or sometimes keypads, or maybe if they're advanced, the fingerprint system.
[06:45.560 --> 06:47.120]  But you could also use face scanning for that.
[06:47.240 --> 06:48.200]  So that's a reverse thing.
[06:48.280 --> 06:51.360]  Your database has the good guys in it, and you're looking for them.
[06:51.360 --> 06:55.600]  And another one is background checks for new hires.
[06:56.720 --> 07:04.220]  The airports now, even though none of the hijackers are involved, you know, work at the airports, there's a lot of concern now about employees at airports.
[07:04.220 --> 07:09.740]  They have access to a lot of equipment in the airlines, and we need to do background checks on them.
[07:09.820 --> 07:15.980]  So as a minimum, the government's forcing all airports to run fingerprint checks on people.
[07:16.180 --> 07:19.780]  But another thing you can do is a face scan if you have some kind of database of bad people.
[07:23.510 --> 07:28.130]  One of the questions is, well, how does a face scanning system work?
[07:28.310 --> 07:29.870]  And it's actually pretty straightforward.
[07:30.210 --> 07:34.850]  At least for this crowd, I don't think you have too much trouble understanding the basic idea.
[07:35.630 --> 07:39.930]  But you really have two basic components.
[07:40.170 --> 07:49.430]  You have a video camera, and you have a computer with a video converter on it that digitizes it, or you have a digital camera to begin with.
[07:49.430 --> 08:00.090]  So basically, you have a video camera set up at some location that can get good shots of people's faces, you know, pointing at the right level for face shots.
[08:00.530 --> 08:08.390]  And that goes into, it's fed into the computer, converted into a digital format.
[08:08.950 --> 08:15.110]  And then, what the face scanning software does, is it looks for the eyes on the picture.
[08:15.270 --> 08:19.450]  So first it spots a face, and uses algorithms for locating where the face is.
[08:19.690 --> 08:24.590]  And then within that square, it looks for the two eyes, and uses that as sort of an anchor point.
[08:25.210 --> 08:32.050]  And then, it recognizes other features on the face, you know, nose, where your mouth is, and cheeks, and so on.
[08:32.230 --> 08:33.890]  And then takes various measurements.
[08:35.470 --> 08:42.550]  And from this, it creates a digital template, which is, you know, sort of a, kind of like a long number, which represents who you are.
[08:43.670 --> 08:53.370]  You know, each time you take a picture, you're going to get a little bit different number for there, but four people, you tend to end up with the same sort of digital template all the time.
[08:54.030 --> 09:04.330]  And then the second component, is you have a database of pictures in it, of photos, that you've somehow taken at some other previous point in time.
[09:04.630 --> 09:12.010]  And then you have an algorithm that takes that template that you've just measured for the picture you've taken, and then you match against each person in the database.
[09:12.790 --> 09:16.990]  And then you come up with a list of pictures that seem to be close to that person.
[09:17.710 --> 09:22.810]  And if you have a, if you're looking for the bad guys, it sounds the alarm and says, wait a minute, we've got a bad guy here.
[09:23.230 --> 09:29.710]  If you've got a good guide system, you know, where you're letting access into, say, an enclosed area, it unlocks the door.
[09:30.190 --> 09:31.490]  So it can go either direction.
[09:31.590 --> 09:34.630]  But the whole idea is you alarm if you get a close enough match.
[09:35.130 --> 09:35.410]  Yes?
[09:36.430 --> 09:38.390]  Does the angle change?
[09:39.110 --> 09:41.490]  Yeah, well, we'll talk about that.
[09:41.610 --> 09:43.130]  I have some real pictures here.
[09:43.330 --> 09:45.610]  But, yeah, these guys have thought about it, though, something like it.
[09:46.530 --> 09:48.750]  It's one of the problems of doing these systems.
[09:49.050 --> 09:49.110]  Okay.
[09:51.610 --> 09:54.130]  So, but we have this concept of the alarm.
[09:54.310 --> 09:58.290]  And the alarm can either be a good alarm or a bad alarm, depending on the system that we're trying to build.
[10:03.700 --> 10:05.500]  Now we get into the matching system.
[10:05.780 --> 10:09.180]  And I actually played around with the Visionics face-it software.
[10:10.060 --> 10:12.720]  And, you know, each software package works a little differently.
[10:12.860 --> 10:15.400]  But this is the same idea across the different vendors.
[10:16.560 --> 10:22.700]  If we have two photos that we want to measure, we have the two templates where we've taken all these measurements.
[10:22.700 --> 10:29.860]  And if I've got it right, I think templates are about 80 bits, or it might be 80 bytes, I think.
[10:30.620 --> 10:34.800]  But what we do is we have an algorithm that matches two templates together.
[10:35.020 --> 10:38.440]  And then it rates the match on a scale from 0 to 10.
[10:38.680 --> 10:43.400]  Where 0 means no match whatsoever, and 10 means identical match.
[10:45.800 --> 10:58.040]  And we then have a threshold value where we say if the matching value is greater than or equal to the threshold, then we probably, these two pictures represent the same person.
[10:58.700 --> 11:02.660]  And in the case of Visionics, they have a default value of 8.5.
[11:04.140 --> 11:05.600]  However, you can adjust that.
[11:05.740 --> 11:07.920]  You can adjust that value one way or the other.
[11:08.340 --> 11:09.520]  I do have a quick question here.
[11:09.580 --> 11:11.580]  I was actually going to start off with this, but now I just remembered this.
[11:12.260 --> 11:14.540]  Do we have any Law & Order fans here in the crowd?
[11:15.060 --> 11:16.340]  Just show of hands, some TV program.
[11:17.220 --> 11:23.660]  Criminal Intent had a show a few weeks back here, or at least it was on, I taped it, where they used face recognition systems.
[11:23.820 --> 11:24.520]  Did you see this show?
[11:24.680 --> 11:29.340]  Where they had this, this is the hacker guy who was stalking the woman.
[11:29.580 --> 11:31.060]  He put video cameras in her.
[11:32.460 --> 11:32.800]  Yeah.
[11:34.600 --> 11:35.000]  Yeah.
[11:35.720 --> 11:38.300]  But it was, it's really appropriate for this audience.
[11:38.420 --> 11:41.560]  So the hacker guy was stalking this woman in her apartment with video cameras.
[11:43.800 --> 11:47.680]  And the cops discovered the cameras, but didn't know who put them in there.
[11:47.760 --> 11:48.680]  Anyway, they traced it all back.
[11:48.720 --> 11:51.240]  He had a website where he was observing her and all this stuff.
[11:52.180 --> 11:58.420]  But as an ironic moment is the way that they caught him was using a face recognition system.
[11:58.960 --> 12:03.660]  And so it was sort of, you know, he was videotaping, but he got caught by the video system.
[12:04.280 --> 12:11.420]  And in the show, they did a pretty good job on the face recognition part because they said, oh yeah, we cranked the thing down really low.
[12:11.420 --> 12:15.300]  So they ended up having two false positives first.
[12:15.580 --> 12:17.480]  They grabbed two guys who wasn't the guy.
[12:17.620 --> 12:19.900]  And then the third one turned out to be the person.
[12:20.640 --> 12:24.680]  But it just shows, you know, if they said right in the show, yeah, we set that threshold down really low.
[12:24.680 --> 12:26.260]  Just to make sure we catch them.
[12:32.060 --> 12:36.500]  So, you know, there's a little bit of art, if you will, of choosing that right threshold value.
[12:36.580 --> 12:39.680]  If you set it too high, you'll miss the bad guys.
[12:40.560 --> 12:44.940]  Or if you're building a good guy system, you won't let the good guy, you know, in through the door.
[12:45.080 --> 12:45.820]  And that's really bad.
[12:45.960 --> 12:50.220]  You know, if you've got somebody trying to do a job and they can't get through the door to get to their place, that's bad.
[12:50.840 --> 12:54.860]  Also, but, you know, if you've got a bad guy system, that system's in there to catch people.
[12:55.740 --> 13:00.620]  If you set it too low, you get the reverse problem, which is the system cries woof.
[13:00.900 --> 13:05.880]  And it will say two people are the same person when they're really not.
[13:06.080 --> 13:09.000]  And that's what happened there on Law & Order.
[13:09.100 --> 13:11.560]  They had two people that they misidentified.
[13:12.240 --> 13:12.320]  Yeah?
[13:27.780 --> 13:29.680]  I'm sorry, I can't, it's a little hard for me to hear.
[13:29.920 --> 13:30.860]  Can you use the microphone?
[13:31.560 --> 13:33.460]  Please, when you have a question, go up to the mics.
[13:45.470 --> 13:48.730]  Oh, yeah, some systems liveliness, some systems have that.
[13:48.730 --> 13:52.090]  You turn it off because it might be in the crowd.
[13:52.350 --> 13:53.310]  It might be quite inconvenient.
[13:53.610 --> 13:54.050]  Yeah.
[13:54.610 --> 13:56.330]  It's like, you know, you can't get an orthogonal.
[13:56.750 --> 13:57.130]  Right.
[13:57.190 --> 14:00.250]  Most systems I know don't really, don't have that kind of protection.
[14:00.650 --> 14:04.570]  Yeah, because they're just watching a video signal all the time trying to grab the pictures.
[14:05.690 --> 14:05.730]  Yeah.
[14:07.950 --> 14:12.970]  But that's another, we'll get into some of the problems here of making this all work.
[14:12.970 --> 14:23.190]  So if we set the threshold too low, we'll finger innocent people or we'll let the wrong guys through the door if we're building a good guy system.
[14:23.590 --> 14:31.890]  And if everything is just right, then we have, we've got a system where it's not crying wolf and we're always letting the right people in.
[14:32.250 --> 14:35.790]  The trouble is it's really kind of impossible to come up with just right value.
[14:35.790 --> 14:42.810]  So we end up trying to sort of balance out between the too high and the too low depending on our application.
[14:48.240 --> 14:56.880]  Now playing around with the face head software, I found that the matching algorithms were very sensitive.
[14:59.000 --> 15:03.120]  And you had to have really good images for the system to work.
[15:04.320 --> 15:14.140]  So some of the things that would affect the ability to match, one is lighting, that you had to have good lighting when you're taking the picture.
[15:15.040 --> 15:24.440]  Also the background, you wanted like a clean background if there were other objects in the background that messed things up, which is sort of strange because it really comes, focuses in the face, but it still solved problems.
[15:26.140 --> 15:33.960]  Facial position and expression matter, you know, so that as I smile or turn my head that would affect the algorithms.
[15:34.900 --> 15:36.920]  Even just wearing glasses or sunglasses.
[15:37.160 --> 15:38.600]  Sunglasses in particular are really bad.
[15:38.700 --> 15:42.120]  Basically, with the Visionic system, you cannot wear sunglasses.
[15:42.380 --> 15:47.520]  And they've even, I saw a presentation on their system, they even said you have to take sunglasses off.
[15:47.840 --> 15:52.680]  Now at an airport, you can tell people, you know, you're going to have to do that whenever the camera's around.
[15:52.940 --> 15:57.120]  But if you're doing stuff on the sly, say like in Ybor City, you don't maybe have that luxury.
[15:58.520 --> 16:02.640]  You also have camera proposition, you know, somebody already mentioned that, whether it's up or down or whatever.
[16:02.840 --> 16:05.980]  So there's a lot of factors that go in here.
[16:07.140 --> 16:12.860]  When you set up a face recognition system, you basically have to control the environment where the camera's in.
[16:13.980 --> 16:15.840]  And this is very important.
[16:15.920 --> 16:17.720]  This is very important for this to work.
[16:17.820 --> 16:20.600]  You just can't throw a camera up any place and just hope it works.
[16:22.420 --> 16:29.940]  I want to go over here now to some actually, some of the tests that I ran here to just give you some ideas here.
[16:34.630 --> 16:37.430]  This is, these are various pictures that I took.
[16:37.590 --> 16:40.870]  I had like a little database, a bad, I guess this is a bad guy database?
[16:41.230 --> 16:42.690]  I don't know which way and when this is.
[16:42.690 --> 16:53.490]  But I had a database of three pictures of myself and then I would try out various poses and various situations to see how that effect matching.
[16:53.710 --> 16:58.870]  So you can see at the top there, the top two pictures, you know, came in pretty good.
[16:59.010 --> 17:00.530]  7.8 and 8.
[17:00.630 --> 17:06.410]  I actually lowered the lighting and it worked out a little bit better with the database, you know, pictures that I had.
[17:08.630 --> 17:13.330]  You put a hat on, you can see that dropped my thing down to 7.7.
[17:13.550 --> 17:16.110]  And I wore glasses, you know, that's 7.6.
[17:16.830 --> 17:19.450]  And, you know, compared to 7.8.
[17:19.550 --> 17:20.970]  So there's a little bit of effect here.
[17:21.430 --> 17:26.590]  But if we set a threshold of 7.5, we would still catch all these here.
[17:31.270 --> 17:33.810]  Supposedly it doesn't matter, but my wife would kill me if I shaved.
[17:33.810 --> 17:36.850]  I couldn't run that test, you know, my wife just wouldn't.
[17:40.390 --> 17:44.030]  But you can see here, if I look up, if my eyes are looking up, that drops it.
[17:44.110 --> 17:45.790]  If I smile, that drops it.
[17:51.450 --> 17:52.270]  Close the eyes.
[17:52.370 --> 17:55.650]  Remember, eyes are used as the anchor point, so that had a pretty lowering effect.
[17:55.810 --> 18:00.430]  And if I look down, but, you know, you can see really, sunglasses just kill this system.
[18:00.810 --> 18:03.910]  All of a sudden we go from 7.2 to 5.9 here on this.
[18:03.910 --> 18:08.750]  So, you know, this is why Vizionics say you must take sunglasses off for their system to work.
[18:10.010 --> 18:14.770]  But the more interesting cases really are down here as you look away.
[18:15.870 --> 18:18.770]  And, you know, I didn't measure that 20 degree turn.
[18:18.950 --> 18:20.830]  But you can see, I can still see my eyes.
[18:20.950 --> 18:23.410]  And I've never really turned that far away from the camera.
[18:24.290 --> 18:26.490]  And it's just totally messed the system up.
[18:27.530 --> 18:34.090]  So, in a situation where you're looking for the terrorists, you may have an opportunity to look at two or three or four frames.
[18:34.590 --> 18:43.030]  But if they've got their head turned away from the camera at just the right time, then the system is not going to recognize them.
[18:50.850 --> 18:51.910]  I don't know.
[18:52.170 --> 18:56.190]  I'm not an expert at their software, so that would be a good question for them.
[18:56.290 --> 18:58.150]  And I can't answer that question.
[18:58.450 --> 19:07.690]  But what it seemed to me was, what's interesting here is in the database that I have, I had all good pictures of myself.
[19:07.690 --> 19:12.570]  You know, they were posed pictures that I was, you know, like you do for a driver's license or something like that.
[19:12.890 --> 19:16.030]  So, a lot of these same factors kick in on the database photos.
[19:16.230 --> 19:19.030]  You know, what if your head is turned one for the photo in the database?
[19:19.890 --> 19:26.730]  So, what I kind of found is that it did match up better if I had two pictures where they were both turned away.
[19:26.910 --> 19:28.570]  That gave a fairly decent match.
[19:29.170 --> 19:35.170]  But, you know, so if the terrorists is looking this way one time and this way the other time, you know, there would be no match possibilities at all.
[19:35.890 --> 19:36.250]  Yeah.
[19:40.290 --> 19:41.550]  Okay, that's a good question.
[19:42.610 --> 19:44.350]  I bet it would be like the sunglasses.
[19:44.630 --> 19:45.950]  But I didn't think of that one.
[19:46.630 --> 19:46.850]  Yeah.
[19:47.210 --> 19:50.350]  So, I think people would kind of look at you at the airport.
[19:50.530 --> 19:53.390]  But sunglasses, you know, that would be a pretty normal thing.
[19:53.710 --> 19:55.650]  And so they say, you've got to take them off.
[19:55.790 --> 19:55.850]  Yeah.
[19:57.110 --> 20:04.850]  Using the eyes as the basic parameter, a lot of people have very strong corrections that kind of make the eyes smaller, make them further, closer.
[20:04.850 --> 20:08.090]  Is that sufficiently complicated for it?
[20:08.230 --> 20:12.190]  Or would it take off a really strong pair of glasses inside of people?
[20:15.430 --> 20:16.310]  That's a good question.
[20:16.470 --> 20:18.950]  Again, that's another test I made.
[20:19.170 --> 20:23.310]  But, you know, clearly, if you want to put one of these systems in, those are the kinds of tests you need to run.
[20:23.410 --> 20:24.810]  I just didn't think of all those, you know.
[20:25.590 --> 20:28.750]  You know, eye patch or thicker glasses or anything like that.
[20:28.750 --> 20:32.150]  But those are the kinds of tests you'd want to run here for a real installation.
[20:33.850 --> 20:35.870]  Could you guys use the microphone, please?
[20:36.330 --> 20:37.670]  We need to get it recorded.
[20:38.050 --> 20:38.470]  Okay.
[20:38.610 --> 20:40.230]  I'll try to say the questions.
[20:46.520 --> 20:46.960]  Yeah.
[20:46.960 --> 20:53.180]  The question is, is if you use makeup and you put like a fake nose on or something like that, could you fool the system?
[20:54.480 --> 20:55.360]  That may.
[20:55.620 --> 20:57.400]  It depends on if it's something being measured.
[20:57.400 --> 20:58.960]  But I know that facial hair doesn't.
[20:59.040 --> 21:00.460]  They don't really look too much at the mouth.
[21:01.100 --> 21:04.700]  So, you know, as you saw up there, when I smiled, that changed things some, so not a lot.
[21:05.040 --> 21:05.120]  Yeah.
[21:05.680 --> 21:11.640]  I was personally thinking about a hat or a cap that had all these fake eyes over it.
[21:11.720 --> 21:13.440]  Because the system triggers on eyes so much.
[21:13.800 --> 21:14.120]  Yeah.
[21:14.240 --> 21:14.940]  So what is that?
[21:15.020 --> 21:16.680]  That's sort of the rest me on the terrorist hat.
[21:16.800 --> 21:16.980]  Yes.
[21:18.620 --> 21:20.620]  I don't think you need the face recognition for that one.
[21:20.720 --> 21:20.960]  But yeah.
[21:21.500 --> 21:21.820]  There's...
[21:25.450 --> 21:27.230]  Well, version two of the study, I should...
[21:27.230 --> 21:28.070]  I'll have to try all these.
[21:28.070 --> 21:30.410]  I hadn't thought about some of these tests.
[21:31.430 --> 21:31.910]  Yeah.
[21:32.230 --> 21:32.670]  There's...
[21:32.670 --> 21:34.450]  We have a...
[21:34.450 --> 21:38.010]  In Holland, we have an OCR system in place for reading license plates.
[21:38.270 --> 21:41.690]  And the license plates are nice bright yellow with black letters on it.
[21:41.790 --> 21:45.390]  And it's about this big and all the letters are the same font and really wide and nice.
[21:46.150 --> 21:47.070]  So I figured...
[21:47.610 --> 21:55.530]  And they're doing now speed checks where they check speed over a trajectory, checking your license plate in and out of a given range of highway.
[21:56.450 --> 22:08.270]  So I thought of actually devising a company logo for the back of the car, which consists of all these license plates elements sort of splattered in one giant explosion.
[22:08.550 --> 22:09.430]  Because the system...
[22:09.430 --> 22:17.570]  All these systems work by scoring different elements in the frame for what's the chance of this being the eyes as opposed to that, or this as opposed to that.
[22:18.130 --> 22:23.010]  So if those algorithms work, it really helps you to screw them up.
[22:23.190 --> 22:23.350]  Yeah.
[22:23.390 --> 22:24.890]  Well, we don't want to help the terrorists too much.
[22:25.010 --> 22:26.490]  So let's not give too many ideas here.
[22:27.470 --> 22:28.350]  Although we'll get it.
[22:28.410 --> 22:30.270]  I think there's plenty of other problems that we don't have to worry about.
[22:30.410 --> 22:30.430]  Yeah.
[22:30.870 --> 22:32.030]  Maybe it was covered already.
[22:32.130 --> 22:32.790]  I walked in a little late.
[22:32.950 --> 22:35.250]  So forgive me if this was talked about already.
[22:35.690 --> 22:41.530]  But as far as the system goes, does it use one photo per person or is it multiple photos?
[22:41.870 --> 22:43.210]  You can put multiple photos in.
[22:43.370 --> 22:43.570]  Okay.
[22:43.690 --> 22:56.850]  Because I was going to say, you know, would it be possible to say, you know, to devise a system that would take an average of all the keys generated by different pictures, like say, head down, eyes closed, eyes open, and do an average of all the numbers for the photos?
[22:57.070 --> 22:58.670]  There may be that possibility.
[22:58.890 --> 23:03.970]  The system doesn't have this concept of multiple photos per person.
[23:04.190 --> 23:06.770]  You can put them in, but there's no correlation between them.
[23:07.810 --> 23:11.550]  So that's a good question for physionics, but I just don't know the answer for it.
[23:11.810 --> 23:12.090]  Okay.
[23:12.190 --> 23:14.190]  But you can put multiple photos in.
[23:14.490 --> 23:17.990]  And so some photos may track better than other ones.
[23:18.430 --> 23:18.510]  Yeah.
[23:18.510 --> 23:18.610]  Yeah.
[23:18.610 --> 23:23.670]  But I mean, if you have multiple photos of a person in different positions, then it will match them.
[23:24.430 --> 23:24.730]  Yeah.
[23:24.770 --> 23:28.150]  If you're looking for them, the best photo of them will be matched.
[23:28.350 --> 23:28.490]  Right.
[23:28.610 --> 23:28.770]  You're right.
[23:28.890 --> 23:29.050]  Exactly.
[23:32.050 --> 23:32.770]  I'll get into that.
[23:33.030 --> 23:33.050]  Okay.
[23:34.030 --> 23:34.210]  Yep.
[23:36.250 --> 23:40.790]  Everybody's asking sort of the reverse of what I'm interested in, how to fool the system.
[23:41.210 --> 23:52.390]  I'm more interested in when the system is set, if you set the system to 9.9 or something, what is the reliability that when it does hit a match, that it is actually a match.
[23:52.570 --> 23:52.850]  Okay.
[23:54.090 --> 23:56.770]  I'll get into that, but if I don't, remind me.
[23:57.230 --> 23:57.510]  Okay.
[23:57.830 --> 24:00.970]  Because that's part of the sort of the catching thing here.
[24:01.710 --> 24:01.910]  Okay.
[24:02.250 --> 24:10.290]  So, you know, on this part here, you know, the main thing you can see that this is not sort of perfect technology by any stretch of the imagination.
[24:12.590 --> 24:13.630]  Yeah, I can.
[24:13.990 --> 24:21.070]  By the way, you know, I actually showed, did this dog and pony for Boston Logan Airport, who was considering putting this system in.
[24:23.030 --> 24:25.390]  And I only was given 10 minutes.
[24:25.930 --> 24:28.310]  But anyway, one of the ladies said, thank you very much.
[24:28.310 --> 24:38.070]  The way it got me off the stage was, and she says, you know, we really, she says, she kind of implied my pictures were less than flattering, let's say.
[24:39.770 --> 24:41.610]  But it was a good joke at the time.
[24:41.750 --> 24:43.830]  It was a nice lightened up thing here.
[24:45.990 --> 24:48.290]  I think she was thinking of my smile pictures here.
[24:48.770 --> 24:53.390]  But as you can see, just, you know, turning at different places, we get different match levels.
[24:53.790 --> 24:55.790]  And then these are the database pictures here.
[24:57.990 --> 24:59.810]  Some of them, yeah, that one I'm kind of grim.
[25:00.130 --> 25:01.390]  I guess maybe that's the one you didn't know.
[25:06.820 --> 25:08.320]  Yeah, no, I didn't.
[25:08.720 --> 25:11.740]  You know, but those are all good things.
[25:11.820 --> 25:13.660]  It would be interesting to see how they match up.
[25:17.330 --> 25:18.850]  Do you think ears are a big part?
[25:18.990 --> 25:19.990]  Yeah, I don't know.
[25:19.990 --> 25:20.670]  I don't believe so.
[25:20.730 --> 25:21.950]  I don't believe they're part of the algorithm.
[25:34.920 --> 25:36.980]  There was a question about age.
[25:38.200 --> 25:50.080]  And so what I did was I took various pictures that I had, snapshots that I had in the shoebox in my house, scanned them in, and then just did a check here.
[25:51.780 --> 25:58.880]  Now, what it looked like, you know, so I have these pictures here, you know, 83, 89, 93, 96, 98, and 2000.
[25:59.100 --> 26:01.660]  Then I just sort of crossed, did all the cross between them here.
[26:02.760 --> 26:12.980]  And what my feeling was, I'm not a great statistician, and my eyes here are a little bit bad, but it seemed like there were still good matches between different years.
[26:12.980 --> 26:16.840]  But the closer the pictures were in time, the better the matches were.
[26:17.660 --> 26:19.500]  You know, you just sort of look at that top row.
[26:19.660 --> 26:23.240]  You know, you get your best match in 8-1, then drops to 7-5, but then it goes back up.
[26:23.980 --> 26:26.040]  But, you know, there's some exceptions here.
[26:26.220 --> 26:29.860]  But, you know, there is some age effect, but it's not that bad.
[26:30.040 --> 26:33.380]  You know, there's almost a 20-year span on here, and we're getting pretty good match values.
[26:35.160 --> 26:35.520]  Yeah.
[26:35.900 --> 26:36.840]  Yeah, I have a question.
[26:37.080 --> 26:39.980]  Are these systems tuned at a specific ethical group?
[26:41.420 --> 26:42.340]  Not really, no.
[26:42.720 --> 26:43.600]  No, they're just set up.
[26:43.620 --> 26:44.940]  They're just doing matches of templates.
[26:45.280 --> 26:54.160]  So, there may be, there are some people that were, any kind of biometric system works worse than other people.
[26:54.380 --> 26:56.960]  And so there may be ethnic groups that tend to work worse.
[26:57.180 --> 27:02.680]  But pretty much no, there's not really any ethnicity sort of biases built in here, as far as I know.
[27:03.180 --> 27:07.620]  Yeah, you can't, you know, for one thing, there's no way you can say, you know, like, spot all the Arabs for me.
[27:08.620 --> 27:09.900]  Or spot all the Americans.
[27:10.220 --> 27:11.040]  Or spot all the Brits.
[27:11.140 --> 27:12.000]  You can't do that sort of thing.
[27:14.960 --> 27:16.300]  Well, maybe, to some degree.
[27:16.560 --> 27:16.800]  Yeah, I think...
[27:19.460 --> 27:21.500]  Yeah, you could, that you could probably do.
[27:21.580 --> 27:22.500]  Not with this software.
[27:22.660 --> 27:23.680]  You'd have to have different software.
[27:23.860 --> 27:24.520]  But I...
[27:25.960 --> 27:26.320]  Yeah.
[27:26.320 --> 27:30.760]  What are the points they look at and recognize besides the eyes and the eyes?
[27:31.860 --> 27:34.240]  It's just, basically, this area of the face.
[27:34.500 --> 27:34.860]  I...
[27:34.860 --> 27:35.660]  There's a lot of...
[27:35.660 --> 27:37.500]  And there's something they claim up to 80 points.
[27:37.640 --> 27:38.560]  But it's sort of that area.
[27:39.780 --> 27:40.100]  Okay.
[27:43.370 --> 27:45.390]  Just to give you an idea of the pictures here, you know.
[27:46.650 --> 27:49.270]  All these have beards though, so I don't have any non-beard ones.
[27:49.930 --> 27:51.510]  There's a passport photo.
[27:53.170 --> 27:53.810]  Yeah.
[27:54.210 --> 27:54.570]  They...
[27:56.310 --> 28:01.550]  They do say that, yeah, it's after about age 16 that your face stabilizes.
[28:06.650 --> 28:07.010]  Okay.
[28:09.190 --> 28:09.870]  Here's some...
[28:09.870 --> 28:11.870]  The hijacker test page, which is kind of fun.
[28:16.430 --> 28:19.330]  I grabbed these photos off the FBI website.
[28:19.870 --> 28:21.650]  And you can see, we got a pretty good match.
[28:21.770 --> 28:23.510]  8.4 between these two photos.
[28:23.710 --> 28:26.710]  And to my eyes, the one on the left, you know, he gained some weight.
[28:26.890 --> 28:27.950]  You know, he was a good eater there.
[28:29.510 --> 28:31.290]  But 8.4 is a good match.
[28:31.470 --> 28:34.610]  So you can see, just, you know, aging effect doesn't...
[28:34.610 --> 28:35.050]  matter here.
[28:36.290 --> 28:37.050]  Here's another one.
[28:37.430 --> 28:38.670]  Those are very different pictures.
[28:38.790 --> 28:41.170]  I think a lot of people would say, hmm, is that really that same guy?
[28:41.450 --> 28:42.890]  But they came up pretty good there.
[28:44.590 --> 28:47.170]  And, you know, we had lower quality pictures here.
[28:47.270 --> 28:48.810]  Even on the left one, it claims excellent.
[28:48.910 --> 28:50.170]  I don't think anybody would call it excellent.
[28:50.370 --> 28:51.890]  But, you know, that's a fairly low one.
[28:53.370 --> 28:53.730]  Okay.
[28:54.690 --> 28:58.610]  The next thing here is...
[29:00.490 --> 29:03.330]  Well, let's take those two different tests and combine them together.
[29:07.550 --> 29:10.050]  So, match me up against the hijackers here.
[29:11.070 --> 29:13.090]  7.4, that's a pretty good match.
[29:18.810 --> 29:22.530]  You know, what's kind of scary though, is I keep matching up to that same guy here.
[29:24.010 --> 29:25.190]  How does he feel?
[29:25.870 --> 29:26.850]  Yeah, well...
[29:26.850 --> 29:27.590]  He's dead.
[29:31.950 --> 29:37.110]  So, I probably wouldn't set the alarms off here, but, you know, I kind of came kind of close.
[29:37.470 --> 29:38.890]  7.4 is a little closer.
[29:39.030 --> 29:42.190]  You know, I kind of thought 7.5 threshold would be a good one.
[29:42.190 --> 29:45.530]  You know, so I came pretty darn close here on these here.
[29:47.470 --> 29:49.850]  But, these other ones there were not too good at matches.
[29:50.010 --> 29:50.830]  But, I come up...
[29:50.830 --> 29:52.850]  This guy, I've got a lot in common with him.
[29:53.230 --> 29:54.110]  I'm not sure why.
[29:54.750 --> 30:01.270]  All of these values are not scaled according to the size of the database or within a large corpus of other faces.
[30:01.490 --> 30:03.310]  It's all just absolute individual matches.
[30:03.310 --> 30:03.670]  Right.
[30:03.930 --> 30:04.110]  Yes.
[30:04.270 --> 30:04.370]  Right.
[30:05.230 --> 30:13.110]  So, the probability of a mismatch is going to grow, you know, as the size of the database gets bigger, then the probability is a little better and it messes up.
[30:14.070 --> 30:14.210]  Okay.
[30:22.430 --> 30:22.810]  Okay.
[30:24.130 --> 30:25.890]  So, that was...
[30:25.890 --> 30:28.270]  I gave you, you know, what my actual results were there.
[30:28.470 --> 30:29.630]  And these are some of the...
[30:29.630 --> 30:30.930]  Here are some of the...
[30:31.370 --> 30:32.510]  Some of the things I noticed.
[30:33.690 --> 30:35.650]  You know, lighting makes a big difference...
[30:36.050 --> 30:37.810]  Makes a big difference in image quality.
[30:39.570 --> 30:40.410]  A good...
[30:41.030 --> 30:42.390]  Neutral background is good.
[30:42.630 --> 30:43.870]  Sunglasses cooled the system.
[30:44.490 --> 30:46.250]  Which Visionics now says publicly.
[30:46.590 --> 30:47.490]  They didn't used to.
[30:47.650 --> 30:48.490]  It's kind of interesting.
[30:48.830 --> 30:50.090]  They didn't mention that problem.
[30:50.170 --> 30:50.570]  Now they do.
[30:51.950 --> 30:53.770]  Like I said, I got many false negatives.
[30:53.930 --> 30:56.730]  They recommend a default threshold of 8.5.
[30:57.030 --> 30:58.010]  I thought it was way too high.
[30:58.090 --> 30:59.710]  I never even got close to that match.
[31:00.030 --> 31:02.250]  You know, and I thought 7.5 was pretty good.
[31:02.750 --> 31:06.170]  So, if I had 7.5 threshold, I got no real false positives.
[31:06.610 --> 31:08.410]  But I didn't do really extensive testing.
[31:10.310 --> 31:14.450]  And so, for same people, you know, I definitely saw the system kind of working.
[31:14.670 --> 31:15.170]  You know, which is...
[31:15.910 --> 31:17.050]  I guess I did get some...
[31:17.050 --> 31:18.210]  I got 1.9 value.
[31:18.390 --> 31:21.850]  So, for the same person, I got matched values between 7 and 9 for different people.
[31:22.390 --> 31:24.770]  Typically saw, you know, 5.5 to 7.
[31:24.770 --> 31:28.130]  You know, so you can see the system kind of does match up with people.
[31:28.290 --> 31:29.970]  As long as you have good quality images.
[31:38.230 --> 31:44.650]  But let's move on now away from the sort of the techie side of things and start looking at the system issues here.
[31:45.830 --> 31:54.390]  In order for the face recognition system to spot the terrorists at the airport, you're going to have to build a database of the bad guys.
[31:56.750 --> 31:58.810]  So, that's an immediate problem there.
[31:59.030 --> 32:02.530]  If we think about, let's say, Al Qaeda might have...
[32:02.530 --> 32:07.490]  Today, I was just reading in the newspaper, 5,000 members that are still in Pakistan.
[32:11.710 --> 32:15.510]  The question is, where are we going to get the pictures for those people?
[32:16.030 --> 32:19.230]  I mean, they're not exactly lining up to have their pictures taken.
[32:19.630 --> 32:21.450]  You know, we do have a picture of Bin Laden.
[32:21.870 --> 32:24.890]  We have pictures of his lieutenants.
[32:25.330 --> 32:30.390]  But I seriously doubt we have pictures of those other 5,000 members.
[32:30.790 --> 32:37.210]  So, for this system to catch the terrorists at the airport, doesn't seem possible.
[32:37.390 --> 32:39.890]  Because we're just not going to have the pictures for those people.
[32:40.330 --> 32:44.230]  So, this idea that we're going to use these systems to catch terrorists doesn't seem very reasonable.
[32:45.430 --> 32:54.890]  Now, if we go to the September 11th hijackers, they actually knew two of them were suspected terrorists.
[32:55.990 --> 32:57.950]  So, you know, we had a little better luck there.
[32:58.650 --> 33:03.090]  There was a guy named Nawaf Azami and Khalid Alamadar.
[33:04.150 --> 33:07.410]  And they may have... I wrote here that they had photos of one of them.
[33:07.470 --> 33:08.790]  They probably had photos of two of them.
[33:08.790 --> 33:18.730]  So, what happened is, in late 1999, they observed these two guys in Malaysia having a meeting with someone who's known to be a member of Al-Qaeda.
[33:19.710 --> 33:24.430]  And the Malaysian, you know, intelligence folks observed this meeting.
[33:24.590 --> 33:25.350]  They didn't record it.
[33:25.350 --> 33:28.930]  There were probably pictures from that meeting, but not 100% sure.
[33:30.070 --> 33:34.330]  But both hijackers had... they were allowed into this country after that meeting.
[33:35.510 --> 33:38.010]  And they actually entered the country twice.
[33:38.350 --> 33:39.530]  They left and came back again.
[33:40.790 --> 33:46.410]  Now, what's sort of interesting here, if you wanted to find Nawaf,
[33:51.140 --> 33:54.480]  he was pretty easy to find.
[33:54.580 --> 33:56.420]  He was listed on the white pages of the phone book.
[33:58.280 --> 34:02.120]  So, this idea that we need a face recognition says Neketje's guy seems kind of silly.
[34:02.600 --> 34:09.920]  Unfortunately, what happened was the CIA recognized this guy was probably, in their terminology, dirty in the year 2000.
[34:09.920 --> 34:15.980]  They didn't let the FBI and INS know until August of 2001.
[34:16.200 --> 34:19.460]  So, for 18 months, these guys were running around, known, dirty guys.
[34:19.700 --> 34:22.520]  But for some reason, the CIA didn't mention this to anyone.
[34:23.540 --> 34:25.460]  But face scanning might have caught these guys.
[34:25.580 --> 34:26.260]  It was possible.
[34:26.580 --> 34:28.340]  You know, we shouldn't discount that.
[34:28.500 --> 34:32.860]  But there's probably other ways to catch these guys, you know, beforehand.
[34:39.120 --> 34:42.080]  Some things about setting up face scanning systems at the airport.
[34:42.080 --> 34:44.700]  I wrote this before they installed any systems.
[34:44.900 --> 34:46.880]  And it turns out the rules are exactly what they did.
[34:47.360 --> 34:50.180]  You generally can't use existing surveillance cameras.
[34:50.340 --> 34:51.760]  You need to set up new cameras.
[34:52.040 --> 34:56.800]  You've got to have a walkway where the system is, where lighting is strictly controlled.
[34:57.180 --> 34:58.060]  There's no windows.
[34:58.060 --> 35:00.540]  It's the best way to do it than just put in artificial light.
[35:01.520 --> 35:04.500]  And then you force people to walk down a single file line.
[35:04.500 --> 35:10.360]  And the cameras must be kind of at head height.
[35:10.720 --> 35:13.500]  And you sort of walk into a wall where the camera is.
[35:14.320 --> 35:16.560]  And the passengers must look straight.
[35:17.080 --> 35:18.720]  And you've got to remove sunglasses.
[35:19.260 --> 35:25.640]  And then you need a human being that, when you get a hit, to verify, okay, we've got two people here that the computer says match.
[35:25.740 --> 35:26.480]  Do they really match?
[35:26.620 --> 35:27.950]  So you need to really have human backup.
[35:28.820 --> 35:35.950]  Now, I attended an airport security conference two weeks ago, or three weeks ago, where Visionics made a presentation.
[35:35.950 --> 35:43.870]  They did an installation at the Statue of Liberty over a weekend when there were some terrorist, vague terrorist threats against the Statue of Liberty.
[35:45.020 --> 35:48.430]  And the gentleman from Visionics described the set up there.
[35:48.520 --> 35:50.140]  And it was like identical to this list.
[35:50.410 --> 35:53.410]  So I had never even professionally installed one of these systems.
[35:53.460 --> 35:58.540]  But you could just see from just using the software that these systems are, are fairly quirky.
[35:58.850 --> 36:01.350]  And they need to have special care.
[36:01.560 --> 36:03.300]  And this is the same list that they had.
[36:06.170 --> 36:29.290]  Now, when we get into the accuracy question, which was asked earlier, overall, my guess is if you had a terrorist walk through, and you had good pictures of them in the database, that you have about a 50 percent chance of getting a match.
[36:29.670 --> 36:32.570]  The vendors claim more around like 90 percent.
[36:33.590 --> 36:35.390]  But it feels more like 50.
[36:35.550 --> 36:49.870]  And these folks felt at the international biometrics group that even under ideal control conditions, systems succeeded in identifying individuals from getting the database only a little more than half the time.
[36:49.870 --> 36:52.730]  So vendors are claiming like 90 percent.
[36:53.470 --> 36:56.130]  Independent testers seem to end up around 50 percent.
[36:56.350 --> 37:00.030]  But that is predicated on one very important thing.
[37:00.510 --> 37:06.410]  You can control everything at the airport, but you've got to have good pictures of the terrorist.
[37:06.650 --> 37:08.310]  And you don't control that situation.
[37:09.110 --> 37:15.470]  You know, even if you've got a picture of the terrorist, that doesn't mean it's going to be up to the standards of being able to put in that database and work well.
[37:15.470 --> 37:17.730]  So that's why this is iffy.
[37:18.270 --> 37:25.150]  Now see, if you're doing a good guy system where you're doing access control, where you allow somebody to go through the door, you're enrolling them into the system.
[37:26.070 --> 37:28.890]  And you control both sets of photos.
[37:29.190 --> 37:31.210]  You know, when they're at the door as well as when they're enrolled.
[37:31.310 --> 37:32.430]  So it's a different, different problem.
[37:32.630 --> 37:35.490]  But when you're looking for terrorists, you don't control those pictures in the database.
[37:36.130 --> 37:41.990]  Now what's extremely interesting is there's been about a half a dozen tests of these systems at airports.
[37:41.990 --> 37:47.910]  I actually got scanned at Boston Logan Airport because they set one up for the Winter Olympics.
[37:48.090 --> 37:50.990]  So I was flying out to Salt Lake and I got scanned there.
[37:53.790 --> 37:58.830]  The way they test these systems is they put into their quote-unquote bad guy database.
[37:59.010 --> 38:01.130]  They create some kind of bad guy database, which I understand.
[38:01.330 --> 38:03.930]  They just download pictures from the FBI website like I did.
[38:04.070 --> 38:07.070]  So they're relatively small, a few hundred photos.
[38:07.070 --> 38:11.930]  But then they take and put employees of the airport into the database also.
[38:12.290 --> 38:15.390]  They're not actually trying to, with these test systems, they're not actually trying to catch anybody.
[38:15.550 --> 38:16.350]  They just keep statistics.
[38:17.010 --> 38:23.790]  And so they put in pictures of employees in these databases and then they measure, you know, they count how many times they recognize employees.
[38:24.250 --> 38:27.750]  But of course those pictures of employees are good, controlled pictures.
[38:27.930 --> 38:30.030]  They're not what you would have for terrorists.
[38:30.030 --> 38:33.850]  So they claim all these great, you know, success rates of 80 or 90 percent.
[38:34.050 --> 38:36.710]  But it's kind of cooked with cookbooks in my opinion.
[38:37.150 --> 38:45.890]  And I was talking with someone who actually is supervising some of this stuff, who works for the TSA, the group that's looking at putting these systems in.
[38:46.050 --> 38:57.610]  And he says, oh yeah, down at St. Petersburg they put in the cops into the database and they took their pictures exactly at the same place with the same cameras that they use for scanning.
[38:57.610 --> 39:00.090]  And that's just not going to be the way it is for terrorists.
[39:00.430 --> 39:03.650]  So one needs to be very skeptical when you hear about these systems.
[39:26.370 --> 39:26.810]  Right.
[39:27.210 --> 39:29.870]  And you kind of stole my thunder here.
[39:30.310 --> 39:30.750]  Thanks.
[39:32.570 --> 39:34.310]  Yeah, but that brings up...
[39:34.310 --> 39:35.650]  Yeah, that brings up the...
[39:35.650 --> 39:42.170]  And that's exactly the point here, is if you start looking at all this, you know, we maybe install these systems in airports.
[39:42.990 --> 39:46.390]  And we can't catch terrorists with them, but that same thing goes.
[39:46.550 --> 39:49.930]  We definitely have mug shots of the local bad guys.
[39:50.130 --> 39:51.630]  So that's who we'll put into the database.
[39:52.210 --> 40:02.070]  And it'll be just like the Super Bowl where we have, you know, we just simply use the airports as a convenient place to round up the usual suspects.
[40:02.810 --> 40:07.390]  And this is exactly the point that I made at Logan Airport when I presented it to them.
[40:07.550 --> 40:12.410]  I said, is Logan Airport the right place to be, you know, rounding up the local bank robbers?
[40:12.410 --> 40:14.310]  You have enough problems with security already.
[40:14.570 --> 40:17.470]  Do you want to take on a burden of the local police department?
[40:18.190 --> 40:23.110]  And the answer, they said, was the head of Logan Airport says, yeah, we don't care.
[40:23.410 --> 40:25.050]  You know, that sounds fine to us.
[40:25.530 --> 40:37.090]  He says, the reason that we want these systems in is for a deterrent effect, that the next time that there's a hijacking, it isn't done from Boston Airport, or because the bad guys think we have this system that'll catch them.
[40:38.050 --> 40:47.570]  And that seems to be... and the same sentiment was echoed at this airport security conference that I went to, that it's more to scare people than anything else.
[40:48.270 --> 40:51.490]  If the bad guys think it's going to work, then they'll stay away.
[40:51.810 --> 40:54.770]  And they may have... there may be something to that.
[40:54.830 --> 40:56.810]  I don't want to say that it's totally bogus.
[40:57.010 --> 41:02.490]  But what I do think is these systems, as far as catching terrorists go, probably aren't going to work.
[41:02.610 --> 41:11.510]  And it's going to be much more, let's just local bank robbers or deadbeat dads or whatever we decide as criminals who we're going to get with these systems.
[41:12.050 --> 41:13.310]  Yeah, I don't know.
[41:18.880 --> 41:19.840]  Technical difficulties.
[41:22.400 --> 41:25.860]  Supposedly casinos were using this a lot in years past.
[41:26.080 --> 41:26.220]  Yeah.
[41:26.260 --> 41:27.580]  What's their experience with this?
[41:27.800 --> 41:29.660]  Especially in Las Vegas, I saw a special one.
[41:30.040 --> 41:30.100]  Yeah.
[41:30.120 --> 41:32.280]  I said they're using this to catch cheaters.
[41:32.980 --> 41:33.460]  Great.
[41:33.860 --> 41:36.280]  Another guy spoiled my thing.
[41:36.480 --> 41:36.500]  No.
[41:36.980 --> 41:38.560]  That was perfect timing on that.
[41:40.820 --> 41:47.000]  I think the casinos, if you look at that application, it really shows you how technologies find their ecological niche.
[41:48.280 --> 41:53.120]  That face recognition has all these sorts of problems or limitations.
[41:54.080 --> 42:00.540]  So the casinos, they bought into the idea of face scanning to look for known cheats.
[42:00.980 --> 42:05.600]  So what they do is the casinos get together and they build a database of all the bad guys.
[42:05.720 --> 42:06.880]  Well, I mean, sorry, the cheats.
[42:07.120 --> 42:10.000]  And I'll put that in quotes and I'll tell you why in a minute why I put it in quotes.
[42:10.420 --> 42:13.080]  And they have companies who run these systems for them.
[42:14.660 --> 42:19.400]  So for them, you know, a casino might pay $200,000 a year to run this service.
[42:19.400 --> 42:26.200]  But as long as it stops a million dollars worth of cheating, it's paid for itself.
[42:26.420 --> 42:28.420]  The ROI on that is quite good.
[42:28.900 --> 42:31.880]  And so for their standpoint, they're just working off the statistics.
[42:32.520 --> 42:39.880]  So, you know, even if this system only gets 10% of the cheats that walk in, it could be higher because they do control the pictures of when people go in the database.
[42:39.880 --> 42:40.820]  It should be higher.
[42:41.040 --> 42:43.380]  It's just a good return on investment.
[42:43.860 --> 42:48.180]  Now, what's interesting about those systems is, well, like, well, who goes on that database?
[42:50.020 --> 42:51.840]  And, you know, who, like, who decides?
[42:52.020 --> 43:01.320]  Well, you know, certainly if you've got somebody who's stealing chips from the table or, you know, you watch the Discovery Channel and they show you all those, you know, physical cheats where they grab stuff.
[43:01.440 --> 43:02.260]  You know, those are crooks.
[43:02.700 --> 43:05.360]  You know, so we could say, well, those are okay that we put in.
[43:05.620 --> 43:08.280]  But they also consider card counters crooks.
[43:08.920 --> 43:13.340]  And card counters are simply people who know how to play blackjack better than the casino does.
[43:14.160 --> 43:19.500]  And they're labeled by the casino as crooks and they get, you know, excluded from going in and playing blackjack.
[43:19.780 --> 43:29.860]  And so you can really see, you know, sort of the, sort of, the, sort of, say, civil rights issue side of this thing here is left up to the casinos.
[43:30.140 --> 43:33.020]  And, you know, they're excluding people that I don't think should be excluded.
[43:33.200 --> 43:41.780]  You know, just because you have to memorize cards in your head and you've learned how to sort of calculate some odds in your head and you know when the right time to bet and not bet, that doesn't seem like cheating to me.
[43:42.820 --> 43:45.980]  So, but anyway, they're using these systems to keep those kind of people out.
[43:46.460 --> 43:46.520]  Yeah.
[43:48.060 --> 43:51.220]  As a civil rights attorney, I'm concerned about the misuse of the system.
[43:51.400 --> 43:58.500]  And I was just wondering if they're going to put white collar criminals into the database, but a lot of CEOs not be able to take airplanes anymore.
[44:01.140 --> 44:03.420]  Well, that would be a great place to round them up.
[44:04.100 --> 44:06.300]  You know, because that's who fly, you know, business people fly.
[44:07.200 --> 44:14.940]  But, but, but no, that's the good point, is these systems, um, fundamentally it gets down to who gets to make that decision and who goes in the database.
[44:15.260 --> 44:18.040]  You know, what's the, the legal standard for putting people in there.
[44:18.360 --> 44:24.100]  You know, mostly what they talk about is people who are wanted, you know, for, you know, jumping bail.
[44:24.100 --> 44:26.600]  Or there's an outstanding arrest warrant for them.
[44:27.180 --> 44:32.440]  And it's not, you know, with the exception of one thing I heard in Ybor City, it's not for people just known to be bad.
[44:32.700 --> 44:35.180]  But it's people who've actually been, you know, they're looking for.
[44:35.180 --> 44:38.040]  But also it opens up a door for misuse.
[44:38.260 --> 44:42.420]  I have personal experience where they didn't use, they just used videotapes.
[44:42.620 --> 44:45.000]  There was a demonstration in front of City Hall in New York.
[44:46.260 --> 44:47.880]  In front of City Hall in New York.
[44:48.060 --> 44:50.920]  And they used videos of a past demonstration.
[44:51.240 --> 44:54.640]  And they thought they caught someone that was supposedly beating up on a cop.
[44:55.440 --> 44:57.480]  Fortunately, the demonstrators got hold of me.
[44:57.580 --> 44:59.320]  I'm with the Mass Defense Committee of the National Lawyers Guild.
[44:59.320 --> 45:01.340]  Went down to the police station.
[45:01.980 --> 45:05.980]  And the two cops that were holding him admitted that he didn't even look like the guy in the video.
[45:06.380 --> 45:07.880]  But they still put him through the system.
[45:08.020 --> 45:08.880]  And we did a police lineup.
[45:09.220 --> 45:10.560]  Unfortunately, he wasn't recognized.
[45:11.100 --> 45:13.740]  But this could be really used just to harass people.
[45:13.900 --> 45:16.700]  To stop them for like hours missing important flights, whatever.
[45:17.140 --> 45:20.700]  And it's, to me, it's very scary that it's another techno fix.
[45:20.880 --> 45:24.460]  And, you know, we have to really be wary of misuse of...
[45:24.460 --> 45:24.480]  Right.
[45:24.680 --> 45:24.920]  Yeah.
[45:24.920 --> 45:26.700]  There's no doubt about these systems.
[45:27.300 --> 45:34.240]  Depending on what level they get put into, you know, can have some very, very, you know, really bad effects.
[45:34.660 --> 45:36.880]  I mean, that's what my concern is.
[45:36.980 --> 45:41.440]  And they're going under the guise of, you know, keeping us safe from the terrorists.
[45:41.580 --> 45:43.220]  But they just don't, they can't be used that way.
[45:43.780 --> 45:46.860]  The other part of this, too, is somebody who travels on airplanes a lot.
[45:47.580 --> 45:51.680]  There's two error rates that we need to talk about when we're talking about these systems.
[45:52.200 --> 45:56.740]  One is the, you miss somebody that's really the bad guy.
[45:57.200 --> 46:01.580]  But then the other reverse is that you finger somebody and say, this is a bad guy.
[46:01.720 --> 46:02.660]  When he really isn't.
[46:03.180 --> 46:06.060]  So the question is, well, what happens in that situation?
[46:06.280 --> 46:07.080]  How do you handle it?
[46:07.160 --> 46:13.640]  Well, the first thing you need to do is you need the human being to go and match up the two pictures and say, oh, yeah, they kind of look the same.
[46:14.000 --> 46:14.960]  Well, then what do you do?
[46:14.960 --> 46:20.880]  You know, if they do, you know, if the human being says they match, then you need to send some security people over and talk to that person.
[46:21.380 --> 46:27.640]  Well, what if he runs or what if he, you know, you know, there's a lot of what ifs there of how you handle that situation.
[46:28.440 --> 46:30.740]  And it can be pretty disruptive at the airport.
[46:30.740 --> 46:39.320]  And you have to consider that the number of terrorists that are going to be going through the airport that are going to be in the database is going to be, you know, one in 500 million maybe.
[46:39.600 --> 46:42.860]  But the number of false positives you're going to get is about one person.
[46:43.380 --> 46:49.600]  The system is, the software is going to say about once every one or two airplanes is the error rates.
[46:49.820 --> 46:52.960]  They're around anywhere as low as a half a percent up to 2%.
[46:53.420 --> 47:00.400]  So you have this whole question of, you know, of, you know, one or two plane loads of somebody being figured as a terrorist.
[47:00.680 --> 47:02.340]  And so you start thinking about that.
[47:02.460 --> 47:05.860]  And what's going to happen is that the security people are just going to give up on this.
[47:05.980 --> 47:09.780]  They're just going to let people go because they're never going to see a terrorist.
[47:09.900 --> 47:12.700]  They're going to keep seeing all these, all the system constantly crying wolf.
[47:13.240 --> 47:16.860]  So there's a lot of sort of very interesting policy decisions that have to be worked out.
[47:17.000 --> 47:19.140]  And this came up at this airplane security conference.
[47:19.140 --> 47:25.400]  People who design security systems for airports in similar kinds of systems have similar problems.
[47:25.560 --> 47:27.700]  So it's not cut and dry how you make these things work.
[47:28.340 --> 47:28.680]  We have two.
[47:29.060 --> 47:36.300]  What kind of studies have they done on twins or family members with obviously similar facial characteristics?
[47:38.360 --> 47:39.860]  I haven't heard.
[47:40.660 --> 47:41.100]  Okay.
[47:41.640 --> 47:42.700]  I'm sorry on that one.
[47:42.780 --> 47:43.980]  But that's another good one to run there.
[47:44.100 --> 47:45.000]  I just need to find a twin.
[47:45.420 --> 47:46.680]  Two questions back here.
[47:46.820 --> 47:47.120]  Okay.
[47:48.720 --> 47:52.160]  How susceptible is the system to disguise and other countermeasures?
[47:53.120 --> 47:57.820]  Well, the clear one is, you know, sunglasses, you know, kills it.
[47:57.900 --> 48:00.840]  But if you add a fake beard on or whatever, that shouldn't.
[48:01.140 --> 48:03.020]  It's just a physical area of the face.
[48:03.260 --> 48:05.380]  So, you know, like a fake nose.
[48:05.580 --> 48:07.260]  But I think people would kind of notice it.
[48:09.820 --> 48:11.140]  No, that wouldn't do it.
[48:11.320 --> 48:11.380]  No.
[48:13.460 --> 48:14.460]  One more question.
[48:16.480 --> 48:19.640]  This may be slightly beyond the scope of this discussion.
[48:19.640 --> 48:31.020]  But one of the things that occurred to me when you were talking about casinos is an ongoing theme in American public life is the privatization of previously public spaces like malls.
[48:31.020 --> 48:38.920]  And I'm wondering what, if any, protections are in place for images collected in public spaces or in sort of quasi-public spaces like malls.
[48:39.000 --> 48:40.480]  And where will these images go?
[48:40.620 --> 48:46.420]  I mean, once the system's got a picture of me walking through the airport, what the access chain is behind that?
[48:46.500 --> 48:48.340]  The security company and then...
[48:48.340 --> 48:49.120]  Okay.
[48:49.700 --> 48:57.420]  Well, in the face recognition systems, what Visionics loves to say is, no match, no memory.
[48:57.640 --> 48:59.880]  They automatically discard the image after that.
[49:00.340 --> 49:15.600]  However, you know, for, in general, for CCTV systems or security cameras, in general, it's usually recorded on some kind of, you know, slow VHS tape.
[49:15.600 --> 49:19.140]  You know, they build these special VHS recorders that save that stuff away.
[49:20.140 --> 49:23.660]  In general, you know, it's in the security department that have these tapes.
[49:23.840 --> 49:30.340]  They might recycle them once every few weeks or once every few months, depending on the retention policy of the company.
[49:30.500 --> 49:31.740]  And then after that, they're discarded.
[49:33.120 --> 49:37.600]  So, you know, the bad news, yeah, this stuff, your images are saved a lot.
[49:39.340 --> 49:40.660]  But that's the bad news.
[49:40.740 --> 49:43.720]  The good news is nobody actually almost ever looks at them.
[49:43.720 --> 49:46.700]  No human being actually is looking at them.
[49:47.060 --> 49:53.100]  But that's what's interesting about face recognition is now you have a computer that's actually trying to look at the images.
[49:53.440 --> 49:59.840]  You know, so you can imagine all sorts of Big Brother applications for this stuff way beyond what these companies are currently talking about now.
[50:00.400 --> 50:01.620]  It's very, very interesting.
[50:01.620 --> 50:11.380]  There's an article in the New York Times Sunday Magazine where the gentleman, Joseph Attic, who was president of Vizionic said, Don't worry.
[50:11.600 --> 50:18.400]  We'll never go the really bad route on these systems because I won't write the software to make this happen.
[50:19.500 --> 50:23.520]  And, you know, I thought a level, you know, that was sort of very arrogant.
[50:23.720 --> 50:30.160]  You know, he's sort of saying, you know, I, you know, I'm, you know, I'm the benevolent Big Brother and I'm not going to do bad things.
[50:30.280 --> 50:30.980]  You know, I'm the techie.
[50:31.040 --> 50:32.060]  I'm in control here.
[50:32.360 --> 50:33.260]  You know, don't worry.
[50:33.860 --> 50:40.800]  And pretty clearly, there are some very interesting policy decisions that need to be made about these systems.
[50:42.280 --> 50:45.920]  And this time, you know, I'm skeptical about whether we can do that.
[50:46.060 --> 50:48.040]  And we seem to be heading more towards a lot more surveillance.
[50:48.360 --> 50:53.060]  But there's plenty of, you know, problems with just general surveillance cameras, you know, today.
[50:54.020 --> 50:54.460]  So...
[50:54.460 --> 50:58.700]  Okay, we've got five minutes, so we've got time for a few more questions.
[51:15.290 --> 51:23.390]  The face recognition that you see here was developed by the United States government through a program called FERET, F-E-R-E-T, over the early 90s.
[51:23.650 --> 51:32.230]  They subsidized the research that went into both Visionics and Visagia systems, and then asked them to bring product to market, and then proceeded to test the product.
[51:32.350 --> 51:34.130]  So it's been going on since 93.
[51:34.290 --> 51:35.910]  I don't know, I'm not a conspirator or anything.
[51:36.070 --> 51:37.050]  That was just the way it happened.
[51:37.050 --> 51:43.270]  And they did FRVT, Face Recognition Vendor Test 2000.
[51:43.650 --> 51:53.110]  And the PDF for the results, this 16 meg of results, the charts of the zoom, receiver operating characteristics.
[51:53.950 --> 51:57.380]  Oh, okay.
[51:57.660 --> 51:58.160]  I'm sorry.
[51:58.520 --> 51:58.820]  Okay.
[51:59.540 --> 52:05.480]  And they just, there were only five vendors in the original one, and all but three dropped out.
[52:06.320 --> 52:07.520]  They're running that again.
[52:07.820 --> 52:08.980]  They've finished the tests.
[52:09.260 --> 52:11.120]  The results will be out on the website.
[52:11.420 --> 52:13.160]  This is the, NIST does this.
[52:13.580 --> 52:15.080]  And you'll be able to see that.
[52:15.240 --> 52:19.880]  This year they had, I think, at least 15 different companies come in to have their stuff tested.
[52:21.920 --> 52:31.300]  The, the, the two companies, the CEO of the two companies, Attic and Colosini, Colosini, Colosini, Colosini, Colosini.
[52:31.320 --> 52:32.060]  Oh, he retired.
[52:32.300 --> 52:32.940]  He retired.
[52:33.520 --> 52:36.740]  They, they did not behave well as businessmen during this period.
[52:36.940 --> 52:38.800]  They hyped their technology terribly.
[52:38.800 --> 52:40.540]  And then they won't took profits.
[52:40.940 --> 52:44.040]  And so it was a very unfortunate circumstance.
[52:44.800 --> 52:45.220]  And...
[52:45.940 --> 52:52.040]  The, the NIST tests are very interesting because this is an attempt to do a systematic study of these systems.
[52:53.820 --> 52:59.040]  Kind of along the lines that I was doing there with looking at various, you know, false positive and false negative rates.
[52:59.280 --> 53:01.400]  And the effects of lighting and these sorts of things.
[53:02.180 --> 53:07.940]  Now, what's going to be interesting is when the vendors, when the results come out, they're going to continue hyping those results.
[53:08.460 --> 53:15.100]  And they're going to use the NIST tests as, as, as ammo to say, hey, we're ready to go into airports now.
[53:15.600 --> 53:16.860]  I know that's going to happen.
[53:17.660 --> 53:28.700]  Why I think the NIST tests are probably going to have, are not going to be applicable for the terrorist application is, they're not really look, they're not really getting into this question of how many pictures of terrorists do we have and what are the quality of those pictures.
[53:46.190 --> 53:49.550]  Yeah, they're sending that threshold really, really low.
[53:50.390 --> 53:52.030]  And that's, that's like in law and order.
[53:52.210 --> 53:54.890]  But, yeah.
[53:57.430 --> 53:58.730]  I've seen also U.S.
[53:58.990 --> 53:59.730]  Army tests that was done.
[53:59.830 --> 54:01.150]  This was done more informally.
[54:01.350 --> 54:04.230]  And they were sort of similar, similar kinds of results here.
[54:04.470 --> 54:06.710]  And that, that's the game you have to, have to play.
[54:06.710 --> 54:09.110]  So I got time for one more question here.
[54:09.690 --> 54:09.810]  Yeah.
[54:15.490 --> 54:16.010]  Okay.
[54:16.110 --> 54:18.730]  I don't know if you had already mentioned this because I came halfway through it.
[54:18.810 --> 54:26.990]  But I do remember reading something in the New York Times about England, specifically London for the most part.
[54:27.150 --> 54:30.890]  And talking about their implementation of cameras all over the city.
[54:30.890 --> 54:35.230]  And they, I guess they've been doing it for quite a few years now.
[54:35.370 --> 54:40.090]  At least three years because they're really afraid of terrorism and all their problems with Ireland.
[54:40.390 --> 54:49.890]  And one thing that they said is that in three years, the only people they've ever caught are, you know, known criminals who have done little petty burglaries.
[54:50.410 --> 54:52.370]  And what they were really looking for is terrorists.
[54:52.630 --> 54:58.890]  But almost all the terrorists that, you know, that were doing suicide bombings, it was their first crime.
[55:01.090 --> 55:03.650]  So in three years, they had basically caught nobody.
[55:04.130 --> 55:04.150]  Yeah.
[55:04.670 --> 55:06.070]  I mean, that's a general...
[55:06.710 --> 55:10.470]  The UK has been on the vanguard of installing security cameras.
[55:10.570 --> 55:14.510]  It was in the early nineties and it was related to IRA bombings.
[55:15.370 --> 55:19.670]  And there is a face-scanning system in one of the town squares in England.
[55:19.770 --> 55:26.850]  But they're mostly just installing regular old security, sort of dumb security cameras, if you will, rather than smart security cameras like this.
[55:27.250 --> 55:28.710]  There is one place where they've done it.
[55:29.530 --> 55:32.070]  Visionics love to tout how crime went down in that area.
[55:32.230 --> 55:33.030]  It's unclear why.
[55:33.730 --> 55:36.470]  They never caught anybody, you know, with the system.
[55:36.730 --> 55:38.250]  So, okay, well, thank you very much.
[55:38.530 --> 55:40.790]  Again, my website is ComputerBytesMan.
[55:40.930 --> 55:42.970]  I do have more information up there on face-scanning.
[55:42.970 --> 55:44.550]  Also be good for questions.
[55:44.850 --> 55:45.030]  Thanks.