Re: Face-Recognition Technology Improves
On Sat, 15 Mar 2003, Bill Stewart wrote: They're probably not independent, but they'll be influenced by lighting, precise viewing angles, etc., so they're probably nowhere near 100% correlated either. I notice the systems mentioned in the study rely on biometrics extracted from flat images. Recent crop of systems actually scan the face geometry by using patterned light (apparently, cheaper than using a laser scanner), resulting in a much richer and standartized (lighting and facial orientation is irrelevant) biometric fingerprint. There's a world of difference between a line of people each slowly stepping through the gate past a sensor in roughly aligned orientation and a fixed-orientation no-zoom low-resolution camera looking at a group of freely behaving subjects at varying illumination. Even with basically single-source nonintegrative biometrics one could do a lot with hi-res camera with zoom actively tracking a single person at a time, using a NIR (skin is far more transparent to IR, resulting in a far richer pigmentation pattern fingerprint to be acquired) for illumination. Then there's gait, a physical body model, etc. Shortwave SAR (SAR for THz wavelenths seems to be doable according to recent publications), so reading body geometry would appear possible. Volatile MHC fragment chemosensors are being developed, a hi-tech variant of Stasi's approach with odor samples and canines. (Calibrated sensors, no need for sensor to be exponsed to the scent before, bit vectors never grow stale). By using multichannel, integrative approaches and more sophisticated DSP the error rate can be eventually brought down arbitrarily low, and simultaneously become increasingly hard to falsify. The costs will come down eventually for such integrative telebiometrics systems realtime connected via wireless to be blanket deployable. Unlike a mobile telephone, you can't switch your body off, or leave it at home. It will be interesting to see what will happen politically once the majority of voters will realize they're living in a strictly unilateral version of Brinworld. - The Cryptography Mailing List Unsubscribe by sending unsubscribe cryptography to [EMAIL PROTECTED]
Re: Face-Recognition Technology Improves
At 12:39 PM 03/16/2003 +0100, Eugen Leitl wrote: On Sat, 15 Mar 2003, Bill Stewart wrote: They're probably not independent, but they'll be influenced by lighting, precise viewing angles, etc., so they're probably nowhere near 100% correlated either. I notice the systems mentioned in the study rely on biometrics extracted from flat images. Recent crop of systems actually scan the face geometry by using patterned light (apparently, cheaper than using a laser scanner), resulting in a much richer and standartized (lighting and facial orientation is irrelevant) biometric fingerprint. But there are two sides to the problem - recording the images of the people you're looking for, and viewing the crowd to try to find matches. You're right that airport security gates are probably a pretty good consistent place to view the crowd, but getting the target images is a different problem - some of the Usual Suspects may have police mugshots, but for most of them it's unlikely that you've gotten them to sit down while you take a whole-face geometry scan to get the fingerprint. - The Cryptography Mailing List Unsubscribe by sending unsubscribe cryptography to [EMAIL PROTECTED]
Re: Face-Recognition Technology Improves
On Sun, 16 Mar 2003, Bill Stewart wrote: You're right that airport security gates are probably a pretty good consistent place to view the crowd, but getting the target images is a different problem - some of the Usual Suspects may have police mugshots, but for most of them it's unlikely that you've gotten them to sit down while you take a whole-face geometry scan to get the fingerprint. I think the security-crazed data gatherers would just want to scan biometrics of every single person passing through the metal detector gates, check them against the list of usual suspects, and insert them in realtime into a central database. Where they will remain, for indefinite time, free for any authorized party to do data mining on. Unless explict laws have been passed preventing this very eventuality, and the systems are actually audited that no data is retained beyond what is necessary for processing. - The Cryptography Mailing List Unsubscribe by sending unsubscribe cryptography to [EMAIL PROTECTED]
Re: Face-Recognition Technology Improves
On Sun, 16 Mar 2003, Eugen Leitl wrote: There's a world of difference between a line of people each slowly stepping through the gate past a sensor in roughly aligned orientation and a fixed-orientation no-zoom low-resolution camera looking at a group of freely behaving subjects at varying illumination. The problem is that's exactly the sort of barrier that goes away over time. We face the inevitable advance of Moore's Law. The prices on those cameras are coming down, and the prices of the media to store higher-res images (which plays a major part in how much camera people decide is worth the money) is coming down even more rapidly. Face recognition was something that was beyond our computing abilities for a long time, but the systems are here now and we have to decide how to deal with them - not on the basis of what they are capable of this month, but on the basis of what kind of society they enable in coming decades. Also, face recognition is not like cryptography; you can't make your face sixteen bits longer and stave off advances in computer hardware for another five years. These systems are here now, and they're getting better. Varied lighting, varied perspective, moving faces, pixel counts, etc -- these are all things that make the problem harder, but none of them is going to put it out of reach for more than six months or a year. Five years from now those will be no barrier at all, and the systems they have five years from now will be deployed according to the decisions we make about such systems now. Bear - The Cryptography Mailing List Unsubscribe by sending unsubscribe cryptography to [EMAIL PROTECTED]
Re: Face-Recognition Technology Improves
On Sun, 16 Mar 2003, Bill Stewart wrote: But there are two sides to the problem - recording the images of the people you're looking for, and viewing the crowd to try to find matches. You're right that airport security gates are probably a pretty good consistent place to view the crowd, but getting the target images is a different problem - some of the Usual Suspects may have police mugshots, but for most of them it's unlikely that you've gotten them to sit down while you take a whole-face geometry scan to get the fingerprint. I'm reasonably certain that a 'whole-face geometry scan' is a reasonable thing to expect to be able to extract from six or eight security-gate images. If you've been through the airport four or five times in the last year, and they know whose boarding pass was associated with each image, then they've probably got enough images of your face to construct it without your cooperation. And if they don't do it today, there's no barrier in place preventing them from doing it tomorrow. Five years from now, I bet the cameras and systems will be good enough to make it a one-pass operation. I'd be surprised if they don't then scan routinely as people go through the security booths in airports, and if you've been scanned before they make sure it matches, and if you haven't you now have a scan on file so they can make sure it matches next time. Bear - The Cryptography Mailing List Unsubscribe by sending unsubscribe cryptography to [EMAIL PROTECTED]
Re: Face-Recognition Technology Improves
Derek Atkins [EMAIL PROTECTED] wrote: Were there really 750 Million Passengers flying through ATL? No, 75 million. If you look at my message again I did correctly say 750,000 for the 1% false positive figure, although I did not type a comma to make it easier to read. Therefore, a better question would be how many UNIQUE assengers flew threw ATL, and then take 1% of that True, but to a first approximation most of the 200,000 average passengers per day in ATL will be unique individuals, so the false positive rate over the entire population is a good indicator of the effect of deploying the system in an airport. In any case, unless the individuals who repeatedly are falsely matched against the database stop travelling, they would increase the overall false postive rate by the same amount that repeat passengers who are not falsely matched decrease the overall rate. The more important number in these trials to ask about is the size of the database. A 1% false positive rate on a large population matched against a database of 5 faces is much worse than the same rate against a database of 50. The article mentioned a watch list size of 3000, which seems like a reasonable size for comparison, but the article implies that there were different trials conducted for the study. Without referring to the original report I can't tell if the 1% FP rate was based on that trial or one with a different size database. Taking into account the imprecision inherent in a news article reporting on a large study, all it is safe to say is that when it says only one subject in a 100 the article is saying only while presenting a really horrific scenario for the airport security people if this system is used to screen all the passengers. -- sidney - The Cryptography Mailing List Unsubscribe by sending unsubscribe cryptography to [EMAIL PROTECTED]
Re: Face-Recognition Technology Improves
At 09:01 AM 03/15/2003 -0500, Derek Atkins wrote: Sidney Markowitz [EMAIL PROTECTED] writes: In addition, only one subject in 100 is falsely linked to an image in the data base in the top systems. Wow, 99% accuracy for false positives! That means only a little more than 75 people a year mistakenly detained for questioning in Atlanta HartsField Airport (ATL), and even fewer at the less busy airports (source Airports Council International, 10 Busiest Airports in US by Number of Passengers, 2001). Were there really 750 Million Passengers flying through ATL??? That number seems a bit high... 750,000 * 100 = 75,000,000 usually (:-), which sounds more credible. No idea how many of those are unique passengers, but there are probably a lot of frequent business travellers going through there many times. Also, I'm not convinced that multiple trials for a single individual are independent. Indeed, one could easily assume that multiple trials for a single individual are highly correlated -- if the machine isn't going to recognize the person on the first try it's highly unliklely it will recognize the person on subsequent tries. It's not like there is a positive feedback mechanism. They're probably not independent, but they'll be influenced by lighting, precise viewing angles, etc., so they're probably nowhere near 100% correlated either. There could be some positive feedback, if they keep photographs of near matches. Another mechanism they could use is the set of names of people expected to fly in and out of the airport, but of course that only works for people who use their real names on airline tickets - it's better for tracking Green Party members than for tracking Carlos the Jackal. - The Cryptography Mailing List Unsubscribe by sending unsubscribe cryptography to [EMAIL PROTECTED]
Re: Face-Recognition Technology Improves
Bill Stewart [EMAIL PROTECTED] writes: Were there really 750 Million Passengers flying through ATL??? That number seems a bit high... 750,000 * 100 = 75,000,000 usually (:-), which sounds more credible. No idea how many of those are unique passengers, but there are probably a lot of frequent business travellers going through there many times. Ok Ok ok. I'm sorry for trying to do math on only 6 hours sleep before a flight. I mis-counted 0's. I'm sorry. -derek -- Derek Atkins Computer and Internet Security Consultant [EMAIL PROTECTED] www.ihtfp.com - The Cryptography Mailing List Unsubscribe by sending unsubscribe cryptography to [EMAIL PROTECTED]
Re: Face-Recognition Technology Improves
Sidney Markowitz [EMAIL PROTECTED] writes: In addition, only one subject in 100 is falsely linked to an image in the data base in the top systems. Wow, 99% accuracy for false positives! That means only a little more than 75 people a year mistakenly detained for questioning in Atlanta HartsField Airport (ATL), and even fewer at the less busy airports (source Airports Council International, 10 Busiest Airports in US by Number of Passengers, 2001). Were there really 750 Million Passengers flying through ATL??? That number seems a bit high... Also, I'm not convinced that multiple trials for a single individual are independent. Indeed, one could easily assume that multiple trials for a single individual are highly correlated -- if the machine isn't going to recognize the person on the first try it's highly unliklely it will recognize the person on subsequent tries. It's not like there is a positive feedback mechanism. Therefore, a better question would be how many UNIQUE passengers flew threw ATL, and then take 1% of that for the number of false positives. I think it's safe to assume that the 99% accuracy for false-positives is over the population, not over the number of trials. -- sidney markowitz [EMAIL PROTECTED] -derek -- Derek Atkins Computer and Internet Security Consultant [EMAIL PROTECTED] www.ihtfp.com - The Cryptography Mailing List Unsubscribe by sending unsubscribe cryptography to [EMAIL PROTECTED]
Face-Recognition Technology Improves
http://www.nytimes.com/2003/03/14/technology/14FACE.html?th=pagewanted=printposition=top The New York Times March 14, 2003 Face-Recognition Technology Improves By BARNABY J. FEDER Facial recognition technology has improved substantially since 2000, according to results released yesterday of a benchmark test by four federal government agencies involving systems from 10 companies. The data, which is the latest in a series of biannual tests overseen by the National Institute of Standards and Technology, is expected to encourage government security officers to deploy facial recognition systems in combination with fingerprinting and other biometric systems for applications like verifying that people are who they claim to be and identifying unknown people by comparing them with a database of images. But the report also highlighted continuing shortcomings, like the poor performance of recognition systems in outdoors settings in which even the best systems made correct matches to the database of images just 50 percent of the time. And it cited outcomes that it said needed more research, like the tendency of the systems to identify men better than women and older subjects better than young ones. The report was strictly a technical evaluation and did not discuss any of the privacy or civil rights concerns that have stirred opposition to the technology. Because the results of the different companies are public, the testing is also expected to become a marketing tool for those who did best, including Identix, Cognitec Systems and Eyematic Interfaces. It is expected to be especially helpful to Cognitec, a tiny German company that is not widely known in the United States, and Eyematic, a San Francisco-based company best known for capturing data from traits like facial structures, expressions and gait to create animated entertainment. ``Face recognition had been just a subdiscipline for us,'' said Hartmut Neven, chief technical officer and a founder of Eyematic. He said that domestic security needs had created a marketing opportunity that Eyematic was gearing up to chase. The results were not as positive for Viisage Technology, which had been among the leaders in 2000. Viisage said that the results, that it identified just 64 percent of the test subjects from a database of 37,437 individuals, were at odds with the strong performance it had been having with big customers, like the State of Illinois. While the government test is the largest for such technology, the number of images in the database was far below the 13 million that Viisage deals with for the Illinois Department of Motor Vehicles, where the company says it has picked thousand of individuals seeking multiple licenses under different names. ``We suspect there must have been human or software errors in how our system was interfaced with the test,'' said James Ebzery, senior vice president for sales and marketing for Viisage. While Viisage scrambles to explain its views to customers and chase down any potential problems in the test, it is taking comfort in the tendency of big companies and government agencies to perform their own testing on their own data before selecting Viisage or one of its rivals. The government's benchmarking was performed last summer but the results were not fully tabulated and analyzed until recently. The report singled out a finding that in ``reasonable controlled indoor lighting,'' the best facial recognition systems can correctly verify that a person in a photograph or video image is the same person whose picture is stored in a database 90 percent of the time. In addition, only one subject in 100 is falsely linked to an image in the data base in the top systems. The report also noted that performance has been enhanced by improving technology to rotate images taken at an angle so that the facial recognition software can be applied to a representation of a frontal view. The data examined whether facial recognition systems could help with the so-called watch list challenge, which involves determining if the person photographed is on a list of individuals who are wanted for some reason and then identifying who they are. Cognitec, the leading performer on that test, gained a 77 percent rating but its success rate fell to 56 percent when the watch list grew to 3,000. -- - R. A. Hettinga mailto: [EMAIL PROTECTED] The Internet Bearer Underwriting Corporation http://www.ibuc.com/ 44 Farquhar Street, Boston, MA 02131 USA ... however it may deserve respect for its usefulness and antiquity, [predicting the end of the world] has not been found agreeable to experience. -- Edward Gibbon, 'Decline and Fall of the Roman Empire' - The Cryptography Mailing List Unsubscribe by sending unsubscribe cryptography to [EMAIL PROTECTED]
Re: Face-Recognition Technology Improves
In addition, only one subject in 100 is falsely linked to an image in the data base in the top systems. Wow, 99% accuracy for false positives! That means only a little more than 75 people a year mistakenly detained for questioning in Atlanta HartsField Airport (ATL), and even fewer at the less busy airports (source Airports Council International, 10 Busiest Airports in US by Number of Passengers, 2001). -- sidney markowitz [EMAIL PROTECTED] - The Cryptography Mailing List Unsubscribe by sending unsubscribe cryptography to [EMAIL PROTECTED]