The humans at BuzzFeed like to do things all the way.
Since Amazon's face recognition system Rekognition recognized 28 US congressmen as criminals, they decided to experiment for themselves.
However, BuzzFeed is not testing this system, but a celebrity recognition AI that belongs to Amazon.
Take 507 images of FBI wanted criminals to AI to match.
â–³ Painter Bob Ross and Mike W. Jackson on the wanted list
As a result, guitarists, actors, and painters have successfully matched the wanted criminals with high scores. AI completed a total of 17 matches with a confidence level of 95% or more.
In another test using the NIST face data set, AI also matched the photos of former US Secretary of State Rice and an African-American man after the arrest, with a confidence score of 96%.
You may be witty and you may have discovered that the focus is not on the match between the star and the wanted criminal, but that these matches are not just barely pulling the individual to make up the number, but the AI's confident answer.
Why is this happening? BuzzFeed also wanted to know, so he contacted Amazon.
Amazon: This is not face recognition
The answer may be more surprising than the results of the experiment.
Amazon said, don't look at the celebrity AI is released as an update of Rekognition, the two are actually completely different products.
Celebrity AI is an entertainment product that will appear in places like social networks or funny apps.
So, all it has to do is to give more similar combinations.
For example, Captain Jack of the Pirates of the Caribbean painted very heavy makeup, which was quite different from what Johnny Depp usually looks like.
Even so, the celebrity AI will still give you a high confidence score, telling you that they are the same person.
And the same set of pictures, if fed to Rekognition, the confidence score of the match between the two may be only 50%-60%.
After all, the latter, as a serious face recognition AI, includes law enforcement purposes in its settings.
In order to emphasize the difference between the two, Amazon even stated that celebrity AI is not a face recognition system at all. It is obviously different from Rekognition in the structure of the model.
The difference has not been clearly stated
But the problem is that ordinary people can't feel such a difference.
At least, Amazon's official documents and promotional materials do not reflect the difference between them.
After talking to BuzzFeed, Amazon humans modified the Rekognition file. The revised version shows the structural differences between the two.
An Amazon spokesperson said that the previous article was too silly, and only wanted to attract people to click in, but did not expect users to see clearly.
As mentioned earlier, celebrity AI is aimed at disguising their actors in film and television works and matching their appearances with real looks.
Therefore, if you only match against a group of wanted criminals, of course the match will not produce the correct result.
Amazon also said that everyone does not care about the confidence scores given by celebrity AI, because the scoring rules for real face recognition systems are completely different.
Didn't write clearly?
In this way, the problem comes again.
Joshua Kroll, an academic at the University of California, Berkeley, said that most of the AI ​​systems used in law enforcement will not disclose technical details.
In this case, what the company says is what it says.
Like Amazon, after Rekognition was exposed to "confirming the congressman as a criminal", the company explained to the outside world that more than 95% of the recommendation was confidence.
Amazon has never explained how the confidence score is derived and what it means.
Kroll said that it is Amazon's own choice to cause such a big confusion to the outside world.
If the company is willing to disclose more details, such as how the system was developed or how it operates, there will be fewer doubts and confusions about Amazon's facial recognition technology.
How to interpret the confidence score?
But now, only a 95% recommendation confidence threshold is given. How should law enforcement officials understand this number?
Clare Garvie from the Georgetown Law Center said that if a person sees a 99% confidence level, they will probably think that the result given by the system is 99% certain.
But this is not the case, it is just the confidence of the AI ​​system itself.
How law enforcement officers interpret this number depends on what kind of training they receive.
This interpretation may also affect the direction of case investigation, and humans are likely to place more emphasis on matching results with high confidence scores.
People have all kinds of worries, partly because AI can investigate all kinds of people, but it is difficult for humans to investigate how AI works.
Human beings are bright, AI is dark.
Endless investigation?
In addition, when face recognition is used for law enforcement, once you are photographed by surveillance, you don't know when AI is investigating you.
Amazon said that the purpose of saving images is to maintain and improve the system.
However, no one knows whether the image will be abused.
Breeze Pro Vape 2000Puffs,Breeze Pro Flavors Disposbale Vape,Breeze Pro 2000 Device Disposable,Breeze Vape Device Pro 2000
Shenzhen Zpal Technology Co.,Ltd , https://www.zpalvapes.com