Hacker News Reader

Comment by bsenftner | Hacker News Reader

Are people still thinking a face image can be used to verify age? That's absurd. Former globally leading facial recognition developer here, and the article lightly mentions using a face image and age verification face analysis - that's not age accurate at all. Ask many ethnicities with experience, "age verification" image analysis is so unreliable it is fraud used in this context.

View on HN

Conversely, people in the UK have mentioned that they looked old enough to purchase age-restricted items at physical stores under an "does they look over 25?" protocol and still asked for ID to purchase them.

Can you share more about this please? I work in the industry and would love to know more about your experience with this verification method.

Well, it's not really a verification method, it's the use of age estimation models in a computer vision sense. The problem with age estimation models is they are only better in statistically unreliable ways within controlled ethnic demographics. That word salad means that age recovery trained algorithms have a variance of accuracy that is difficult to reduce, and when successful is only successful on narrow classifications of ethnicity. Part of the issue is ethnicity carries meaningful changes in age representation. Asian, African and several other ethnicity show age later and significantly more subtle than others. Now add in the existence of large demographics of mixed ethnicity, and then add in the issue of the uncontrolled illumination age verification systems are expected to operate... and age verification computer vision is rendered kind of useless. Kind of a joke. Kind of leading one to think anyone trying to sell a solution here could be dumb or a fraud. Might be some new breakthrough, but could it?

I’m not sure - I think between the NIST tracks for age estimation and the work entities have done to gather large, diverse sample sets shows meaningful progress and perhaps real world usage.

Your points above are valid and real concerns, in addition to liveliness. There is work further to be done and improvements to be made. But it seems to me that they are solvable problems.

These datasets are getting granular, monolid vs non, 12+ different ethnicity sub groups and so forth.

Do you not think that with enough data it’s solvable?

The system I worked on had (it's larger now) 100M faces in the training set, and when I was leaving there was a 300M set in the works. We went to lengths to collect and categorize. It's mixed ethnicity that throws a wrench into ethnic categorization. It is ordinary to have people with a half dozen or more racial compositions, and that pretty much wreaks categorization. We (they) also had a a pretty robust liveness detection, and surgical mask detection with a "see through the mask" feature too (available with certain crazy-tech cameras.)

I think it is convenient for the services and probably the regulators to pretend so.

It allows a class of low quality providers to basically trick people into the industry with products that claim but cannot deliver. Then that customer either abandons, or ponys up the real amount of investment necessary for a real solution, and due to being burned already, they are more diligent in their research. Which the industry wants, because too many knock on tech company doors expecting magic genies to grant wishes.

That assumes that the customer actually wants the problem solved, as opposed to doing the bare minimum to be able to say 'we tried' to the regulators. And that works doubly well if the regulators also just want to do the bare minimum so they can say 'we tried' to the politicians, press, and segment of the public that thinks this is a good idea.

That's a very good point: just as theater.

Reply on news.ycombinator.com