The Uses and Limits of Amazon’s “Rekogntion” Facial Recognition Software

 

A new test by the ACLU demonstrates these limitations

 

 

 

 

The American Civil Liberties Union recently tested Amazon’s facial recognition tech — and the results were less than favorable. To test the system’s accuracy, the faces of all 535 members of congress were scanned against 25,000 public mugshots, through Amazon’s open Rekognition API. Although none of the members of Congress were in any of these mugshot lineup, Amazon’s system nevertheless generated 28 false matches. The ACLU claims this raises some particularly serious concerns about Rekognition’s use by law enforcement and in the legal and medical world.

 

“An identification — whether accurate or not — could cost people their freedom or even their lives,” the group said in an accompanying statement. “Congress must take these threats seriously, hit the brakes, and enact a moratorium on law enforcement use of face recognition.” (ACLU)

 

According to The Verge, an “Amazon spokesperson attributed the results to poor calibration.” However this does not necessarily account for the results. Amazon’s system currently operates with the default confidence threshold of just 80 percent. Yet Amazon claims it recommends at the very least a 95 percent threshold for situations such as medicine and law enforcement where relying on a machine to ID someone could cost them their freedom, life, or worse.

 

“While 80% confidence is an acceptable threshold for photos of hot dogs, chairs, animals, or other social media use cases,” the representative said, “it wouldn’t be appropriate for identifying individuals with a reasonable level of certainty.” (ACLU) Even still, the Rekognition suite does nothing to affect that recommendation during the process of setting it up, and there is of course little to nothing to prevent law enforcement agencies from using the default setting of 80 percent.

 

In May of this year, this tech came into the limelight when the ACLU report was able to show the system being in use by a number of LEO agencies including the police of Orlando, Florida. It is sold as a part of Amazon’s Web Services cloud, and is quite inexpensive with a costs as low as less than just 12 dollars a month for the entire department.

 

Furthermore, this test demonstrated a continuing problem of many facial recognition systems, which have  historically had considerably difficulty    in accurately identifying both women and non-white minorities. Of the 28 false matches, 11 involved black members of congress, although they make up just around 20  percent of the whole of congress itself. Some other systems fair even worse. With the system used by the London Metro Police force producing as many as 49 false matches for every legitimate hit, which then necessitates a manual and time and resource consuming search though these false-positives.

Ostensibly, facial recognition IDs would be confirmed through multiple human sources before an arrest would be made, though many say that even checking faces violates privacy rights. Worse still, it is not hard to imagine a situation where an officer sees a false match that leads him to believe the potential arrestee could be armed and dangerous, and also plant ideas about the person before even really investigating, changing the outcome of a routine stop from routine, to possibly violent, even deadly.

Security Technology of South Texas works with analytics and facial recognition video surveillance in its projects, and are experts in integrating, understanding, and sourcing only the best tech to get your job done, at a price you can feel good about. Let us show you the difference between a local, responsive, company that strives for only excellence and client satisfaction versus the kind of experience we have all come to expect from the detached, hard to reach, and inferior service and installations inherent to the juggernauts of the security industry.

Please contact us through email at admin@gostst.com on our website or via phone at  210-446-4863   24/7

Sources: ACLcomU, Verge.com,  Amazon.