Amazon's "Dangerous New Face Recognition Technology" Says ACLU

By John Honovich, Published May 23, 2018, 02:29pm EDT

The ACLU has caused a stir, with a new report Amazon Teams Up With Law Enforcement to Deploy Dangerous New Face Recognition Technology, explaining:

Powered by artificial intelligence, Rekognition can identify, track, and analyze people in real time and recognize up to 100 people in a single image. It can quickly scan information it collects against databases featuring tens of millions of faces, according to Amazon...

With Rekognition, a government can now build a system to automate the identification and tracking of anyone.

Most importantly, the ACLU requested and received 147-pages of information on Amazon's interaction with US government entities. Based on that evidence, our conclusion is that Amazon is having success with greater potential for mugshot comparisons but much more challenges for 'real time' facial recognition. Inside this note, we share our review of that 147-page file, calling out the most notable passages and what that reveals about Amazon's progress, or lack thereof, in facial recognition.

Missing *****

*** **** ********, ***-******* ***** in *** **** ** that ****** ****** ** Simpson ** **** ******:

**** ** **** ** example ** ********** ******, who ***** ** *** on *** ** ******'****************** *********** *******.

** *** *** * user ** *** **********, the **** ** ******** should ** **********. *** if *** *** *** ACLU, *** *** ****** to **** ** **** afraid ** *** ***** liberties ************ ** ***** mismatches, ******, **********, ****** officers *** *** **** man ** *** ** Simpson.

Best *** **** *********

*** **** ********** ******, however, ** *** *********, clearest **** *** *** service, ** ** ******** explains *** ********:

Cheap *** ******* ***** ***********

****** *********** ******* **** photos ** *** ***** as **** ******* **** been ************ ********* *** years. *******, *** **** that ** ** ** inexpensive ** *** *** few ******* ******* *** month (*.*., '$* * month ** ****') ** disprutive economically.

Not ****-**** 

*** ***** ********** ****** makes *****, **** ******'* marketing *** *** ****'* fears, ** **** ***** searches '****** ******* ** less **** ** *******'. While **** ** ********* compared ** ***** **** manually, ** ** **** slow *** *** ********* and ***** ** ****-**** alerting.

Problem ***** ******

********** ****** ******** **** they *** ** *** and *** ******** **** have ***** ****:

Built ** *** ** ******

*** ********** ****** ******** explains *** ** ***** it ** *** ** the*********** ******* / ***:

**** ** *** * normal ****, ** *** employee ******** ** ********** did *********** ** **** and *** **** **** recognition *** ** *** own '**** ****':

**** ******* *** *** success **** **** ***** on *** ***** ****, video **** ******** *****:

Orlando ****-**** ****** ************ *********

**** ***** ** **** impressive ********** *** **** troubling *** ******* ** the ****'* ****** ***** Amazon's ***** ** *******, where *** **** ********:

*********** **** ************ ** now ********* ****** ******* in ****-****, ********* ** ******, ******** *** ******* to ****** *** “****** of ********” ** ******* rolls ** **** “******* all **** *** ****.”

*** ******* ** **** the **** ***** ** Amazon ******** ***** *** [link ** ****** *********], who ** ******** ****** to ****** *** ******* the ****** **** *** following ***** ****:

***** *** **** **** ***** claims, ** *** ********* or ***** ********. *** one, *** ********** ****** **** cites * ~ **-****** processing ****, *** **** our ********** **** ****** cloud *******, **** *** not ***** ** '****-****' (where ****-**** ***** **** a ****** ** ***).

********, ******* ***************** ***** ****** ** the ***:

*** *** *** ***** of ****** *******, ***** having ******* *********** ** Orlando *** ******* ********** any ** *** ****-**** or *** **** *** city **********.

****, *** *** *****, even ** *** *********, given *** ****** ** processing ******** *** *** certain ******** ******** **** would ***** ******** **** at *****, **** ** more **** **** *******.

ACLU ***** ** *********** **** *** ******** ** ****** ******

**** *****, ** ***** the **** *** ****** the ****** ****** ********** these ******** ** *** government *** *** ********* risks ** **** **********. However, *** **** ***** more ******** **** *** trade *********, ****** *********** in *******. ******** ***** overstated ********* ***** * ************ company *****, *** **** accepts **. ***** **** are ****** ******** *** do *** ********** *** the ********** *****, ** maybe **** ***** **** magnify *** ********* ***** and ****** ** *** problem, **** ** *** technology ** *** **** where **** ****** ** to **.

Outlook ***** - ********* ** ********

*** ********** ****** *** case ** * **** good ******* ** * low-cost *** ** *** technology ** ******* **** a ****** ******* ** scanning *** ********. ** expect **** ** ******** to ******.

*******, *** ****-**** '*** over *** ****' *** of ****** *********** ** much **** ***********, ** these ********* **** **********. 

Comments (19)

Give us sex offender databases with booking photos and let's have our school security cameras go to work!

Not sure if you are serious or not but I think that application could be problematic. 

Considerations:

- How often does someone on the sex offender database go to a given school? Guess. The rarer it is, the more likely the ratio of false positives is terrible (i.e., people matched as sex offenders who are not to actual sex offenders).

- The delay in getting back matches from the system. You have to analyze the video stream for faces, get the faces, send them to Amazon, wait for Amazon to return a match. By that time, the person may have moved quite far.

- Then you need an operator who is going to review each of these alerts (which returns to the problem of pt 1 above).

The problem is we don't know how often a sex offender visits a school. (side note: we are discussing a visitor management system that will do the check, but government doesn't more very quickly so I don't know what year this will happen. Additionally, some administrators seem uncomfortable with any type of check). 

I agree and realize that this system isn't ready for most applications, yet. The ACLU is overreacting, for now.   

Curious - what visitor management system is including facial recognition / criminal / sex offender check in their offering? I could see the value, also since you are taking a picture of the person right in front of you, the quality would be better.

The visitor management system I'm most familiar with is School Gate Guardian. 

For clarification, it doesn't use facial recognition, rather subscription-based and end-user maintained databases. Upon entry, your ID is scanned, and it looks for a database match.

It also uses your ID info to print a temporary sticker-badge with the photo from your ID.

Many school districts think a paper sign-in sheet is adequate for visitor tracking. I do not.

EasyLobby also has a sex offender screening service which searches all 50 state databases based on ID. But it doesn't have facial recognition, either.

A few years ago, I did an implementation at a private school with HID Easy Lobby and tied in the sex offender database.  It used the person's name and DOB to match against the registry (we were scanning driver's licenses to register visitors).  I actually got a hit on Day 1 - just a coincidence that a contractor had the same name and DOB as someone on the registry, but was clearly not the same person.

Lynn, thanks!

just a coincidence that a contractor had the same name and DOB as someone on the registry

What a nightmare for that person, not simply that incident but the thought that it might happen regularly for him. 

Thank for such a comprehensive overview John. Very helpful, please keep up the good work. 

This $6 fees is a strange number - seems like Orlando Police obtained this price by using not GovCloud (which is 332.95$ charge) but regular UsWest (Oregon) AWS Recognition service. 

 

Slava, thanks! The $6 claim comes from Washington, not Orlando. I am not sure what Orlando did there.

Related, for those curious on the $332.95 charge, from page 25:

The above says 277,461 images. That sounds like a lot but, for comparison, there are 10x the number of seconds in a month (i.e., 2,628,000 seconds in a month), so even a single camera only sending 1 image every 10 seconds would generate this.

Hi John, you are right.

Yes, 332$ charge seems to be for GovCloud images, while 3$ for regular US West recognition cloud storage and 2.7$ for US West recognition processing (magic 6$ number).

So i was curious how come in report they wrote about 6$ monthly charge. I don't think public sector can use any other cloud except GovCloud. 

And yes, in regards to amount images one need to process. If we want to  process every 1sec IFrame it would cost 3k$ per camera for GovCloud and 2.6k for regular AWS cloud recognition service. This is just a processing fees of recognition service.

One could say that simple optimization strategy would be to use some kind of motion/face detection to minimize face analytics processing. Public cameras in crowded places would need to be processed 24/7 thought.

I don't know about the government cloud options but I think the $332.95 is the one time charge for uploading the original mugshot database.

Yes, you are right John.

332$ charge is for storage for GovCloud since its price is 1.2$ per 1000 image sin comparison to 0.01$ for 1000 images for regular cloud region.

There is no avoiding this as it is our future. AI technology will evolve and become so useful the bad guy will not be able to go anywhere.

Stop freaking out /tinfoil hat.

I just can't wait for my Alexa phone!

Update: Orlando cancels Amazon Rekognition program, capping 15 months of glitches and controversy | Orlando Weekly exclusive, citing upstream bandwidth issues sending video to AWS:

"We haven't even established a stream today. We're talking about more than a year later," Rosa Akhtarkhavari, Orlando's chief information officer, said in an interview with Orlando Weekly during the second pilot. "We have not, today, established a reliable stream." 

Akhtarkhavari, Orlando's technology czar who spearheaded the pilot, said bandwidth issues prevented city staff from running the powerful software in conjunction with any more than one camera — and sometimes that signal wouldn't work as expected. "We've never gotten to the point to test images," she said. 

Read this IPVM report for free.

This article is part of IPVM's 6,651 reports, 895 tests and is only available to members. To get a one-time preview of our work, enter your work email to access the full article.

Already a member? Login here | Join now
Loading Related Reports