7 ms·
The privacy concerns here are huge, but the much more direct issue is from this quote from the article: > the same programs failed to identify darker-skinned w
by pwg 8y ago
The privacy concerns here are huge, but the much more direct issue is from this quote from the article:
> the same programs failed to identify darker-skinned women up to one-third of the time.
A one-third failure rate means that on average, for every three times you try to return home, you are likely to have the system fail to identify you at least once. Being locked out, as a resident, one-third of the number of times you return home, will get very irritating very quickly.
And for the really unlucky, they may seldom have the system recognize them, so they end up locked out a majority of the time.
A 33% recognition failure rate is simply too high to consider door locks to ones home as a good use of facial recognition.
- mc32 8y agoWhy does the system even care about color in the first place? Can't it measure other facial characteristics, measurements?
- Cenk 8y agoProbably a consequence of being mostly developed by and tested on white people Edit: Take a look at this video, it illustrates the problem quite well with a soap dispenser as an example: https://twitter.com/nke_ise/status/897756900753891328 https://twitter.com/nke_ise/status/897756900753891328
- r3bl 8y agoI distinctly remember this example (about HP's laptops) being all over the Internet a decade ago: https://www.laptopmag.com/articles/hp-webcam https://www.laptopmag.com/articles/hp-webcam
- gph 8y agoNot an expert in this area, but I believe it's harder for the software to pick out shadows/shades on a darker skin, and it relies heavily on those shadows/shades to recognize structure.
- ohazi 8y agoIt doesn't care about color. Dark faces reflect less light, making them more challenging to photograph well. If you have less light illuminating a scene, you'll have more noise and less contrast/dynamic range in the image. You need these things to extract features reliably. It's physics, not racial bias.
- Cenk 8y agoYes, but the result – people actually using door locks with a failure rate of 30% for black women — reflects racial bias. If the locks had the same failure for all people they would definitely not be deemed production ready.
- DoofusOfDeath 8y agoI agree, but I guess there could be unintended bias in the product development process. In particular, when using empirical testing to decide if the system is sufficiently reliable, accidentally using an unhelpful sample of faces. Or even earlier, when selecting the camera and lighting technologies to use.
- jfk13 8y agoGiven how hard it is, even with a pretty high-quality camera, to get a photo of our black cat that actually shows her (charming) face effectively, I find this easy to believe.
- Balgair 8y agoIt's mixed. The issue really is in that the testers aren't 'darker' people, though it's not really their fault either. In color film photography they chose the CYM spectral sensitivities to best represent US white women in heavy make-up under stage lights. It was the 1920/30s afterall. The spectral sensitivities haven't changed much since then as a result (an example of the Lock-In effect in full swing). CCD and CMOS camera dopants are chosen to represent the spectral sensitivities that are similar to film, though physics does come charging in on those. To be honest, yes, it's hard to photograph darker skinned people without sufficient full-spectrum light. Darker skin absorbs more radiation at the solar-peak frequency (~550nm, the light green of most plants), as the melanin is evolved to do. That said, if your film/sensor can see far outside of the solar peak, then you'll pick up darker skinned people better. Old-timey tin-type photographs could see well into the UV (notice how there never are any clouds in pictures taken way back when) and darker skinned people showed up just fine in those old photographs as a result. Fun-Fact: the front facing camera on most mobile-phones does not have and IR filter on it, specifically to help with these issues of darker skin. Next time you are near a handle-less toilet with an IR sensor, you can use the front facing camera to see the little IR LEDs beeping to see if you are there. You can also use the front-facing camera to see security lights and what not, typically those are IR ones too. UV light could also be used, but it is more damaging over time, and thus lower frequency IR light is preferred.
- mirimir 8y agoI guess that, with darker skin, there's less accurate data.
- bsenftner 8y agoLead developer of an FR system here: this issue is dynamic range, darker faces, as well as very pale faces as well as any face that the brightest portions of the face are not far from the darker portions of the face, neglecting highlights is difficult for facial recognition simply because there is less distinguishable detail in faces with less visible feature variation. I see the same behavior in very pale Irish, those from Finland, soft featured Asians, darker than coffee Indians, and Africans. Anyone with very little illumination variation, which is necessary to distinguish their features.
- idrios 8y agoAlong the same lines, imagine someone allergic to bees getting stung on the face causing their face to swell, so the system doesn't recognize them and they can't get into their apartment to grab their epipen. This system could literally kill a person
- Balgair 8y agoHeck, imagine something as mundane a haircut, a beard-shave, or a sunburn from a beach trip.
- codedokode 8y agoHair and skin tone is not used in recognition. The points to build the model are taken from eyebrows, eyes, nose and mouth. So as long as you don't have hair covering your eyes it won't affect anything. Also, face recognition technologies are better than described in the comments; you can take a photo of a random person from the street and successfully find their social network account with publicly available sites. Compared to this, recognising just several hundred faces looks like an easy task.
- TeMPOraL 8y agoYup. If the sophisticated facial recognition systems in flagship smartphones can't recognize you with vs. without prescription glasses, I have no hope that whatever el cheapo systems those landlords will be installing can handle a haircut or even bad lighting.
- mftrhu 8y agoHell, even a failure rate of one in a hundred would become unacceptable pretty quickly - going out multiple times a day is not unusual. One in three is beyond ridiculous.
- TheCoelacanth 8y agoBeing locked out of your home one in three times is an eviction. The landlord isn't even making a pretense at honoring the lease at that point.
- cubano 8y agoThat is an almost impossible stat to believe. My last contract I worked developing a biometric system to extract a consistent "headshot" from all the various mugshots from arrest records in the US, and I worked with the Kairos system to extract pixel coordinate data from the images. The API calls gave me a lot more then just pixel information, and one of the data points was skin color and it was correct like 98% of the time. Now, of course, it's possible that posed mugshots deliver a much more consistent image quality then surveillance cams would, but 33% seems suspiciously high to me.
- fwip 8y agoGetting a person's skin color right is not a high enough bar to let them into your apartment. I don't want every white person getting into mine, and I can't imagine they would be fine with every black person getting into theirs.
- jankotek 8y agoSun glasses, haircuts, hats.. 33% failure rate in real life is very plausible. And even 2% rate would get old very fast. That is once a week...
- crooked-v 8y agoI have to wonder what the legalities would be here of having a handyman come and remove the front door lock after being locked out at night or the like.
- bilbo0s 8y agoI'm pretty sure that would be illegal in most places. What you have to do is call the management company. It sucks, but I would just advise darker skinned people, who this is likely to be an issue for, to rent only in places without this sort of system. Being locked out once out of every three times you come home is not really tenable.
- bobthepanda 8y agoBut then when every large complex moves to this all of a sudden darker skinned people have less housing choice and we’re back in 1950. We could just not allow these systems.
- mbrumlow 8y agoSimple problem! Use a rfid chip embedded din their hand! I wish I had this for my house.
- gjs278 8y agoit doesn’t mean that the same person has a one in three chance of not being recognized. it means one in three people won’t be recognized. but the two of three that are recognized will be recognized pretty much every time. and you can always try again if it misses for some reason.