back to article Auditors slam FBI for shoddy testing of facial-recog tech. But no big deal. It only has 641m images on its systems

Three years ago, the US Government Accountability Office (GAO) found that the FBI had very limited metrics about the accuracy of its face-recognition technology. The official auditors thus made six recommendations to improve the tech's accurate and to ensure those using it comply with privacy laws. Now, in a new report, the …

  1. Garymrrsn

    Not Ready For Prime Time

    Facial recognition is an amazing innovation. So is genetic editing. They are both at a similar stage of development in that we know they work, sort of.

    Would law enforcement have the same enthusiasm if it were suggested that they all be genetically enhanced, you know, make bionic people out of them?

    1. teebie

      Re: Not Ready For Prime Time

      We... don't know that facial recognition works at better than 2% accuracy, based on the Met's trials.

      1. T. F. M. Reader Silver badge

        Re: Not Ready For Prime Time

        The "99% accuracy" claimed by the FBI and "98% false positive rate" in the UK trial deployments are, in fact, perfectly consistent. Read up on base rate fallacy.

        This is something that is taught in every beginner's statistics course, but seems to be routinely missed or misunderstood by proponents of "CCTV+AI on every street corner because terrorism". Probably because "who needs statistics if we have AI?"

    2. NoneSuch Silver badge

      Re: Not Ready For Prime Time

      Until facial recognition works, it's just an amazing concept and nothing more.

      Just wait until a bad facial recognition match results in an innocent person being shot dead.

  2. Shadow Systems Silver badge

    Of course you've heard no complaints...

    It's hard to hear anything when you've got your fingers in your ears, are shouting "LA LA LAAA! I can't hear you!", and have your head wedged so far up your ass that you're using your bellybutton as a parascope.

    1. Anonymous Coward
      Anonymous Coward

      Re: Of course you've heard no complaints...

      "that you're using your bellybutton as a parascope."

      I'm trying to decide if you actually meant periscope or a device employed by those searching for ghosts.

      Both are likely to be more accurate than the FBI's facial recognition...

      1. Shadow Systems Silver badge

        Re: Of course you've heard no complaints...

        At the A/C, re my typo.

        I meant periscope. Thank you for the correction.

        1. Anonymous Coward
          Anonymous Coward

          Re: Of course you've heard no complaints...

          It wasn't meant as a correction - while I can see what you meant, I think the actual comment is funnier if the FBI wear their buttholes as a kind of clerical collar and go looking for ghosts...

    2. Vector

      Re: Of course you've heard no complaints...

      OR maybe it was ...

      "...on display in the bottom of a locked filing cabinet stuck in a disused lavatory with a sign on the door saying ‘Beware of the Leopard."

  3. Mike Ozanne

    "To date, we have not had any violations or notice from the public that they feel like their rights are violated,"

    Does that mean they haven't managed to arrest anyone with it yet...

  4. tfewster Silver badge

    " only tested accuracy when a user requested a list of 50 possible matches to a source image..."

    Does that mean what I think it means: It returns 50 possible matches, and 99% of the time one of them is a real match?

    I've abused statistics a few times myself, but my gut would call that 2% accuracy.

  5. SolidSquid

    "and maintains that face scans can enhance public safety if used properly"

    If the FBI has refused to act on several recommendations by the GAO to ensure it doesn't violate people's rights and that it reports data accurately, no amount of rules built around usage will allow it to be used "properly"

    "To date, we have not had any violations or notice from the public that they feel like their rights are violated"

    How many people have been scanned by it vs how many *know* they've been scanned by it? And how many are told they were initially identified with facial recognition if they're arrested? The FBI is notorious for keeping it's methods quiet, and requiring police forces they give access to those methods keep quiet as well. Plus, whatever they do is never going to catch false negatives. Nobody is going to walk into a police station and say "yeah, I was the criminal but your software didn't pick me out", so there's going to be massive biases created there as well (beyond the existing bias against anyone who isn't white caused by sampling issues)

  6. Mike Lewis

    Clark Kent

    I confuse it when I wear my glasses so I have to remove them before my face gets scanned.

  7. Nick Kew Silver badge


    Glad to see you mention metrics. Measuring performance in any complex AI task is something of a challenge, and one that tends to come with a powerful SEP field that means those few in the research community who notice there's an issue are far too maverick to get funding.

    Many years ago I did postdoc research in the complex AI task of computer speech recognition. I was one of the few who looked at how we and every other research team were measuring (and publishing) performance, and our hopelessly meaningless use of concepts like "accuracy", and tried to suggest more meaningful metrics. The basic SEP everyone ignored was that a system that could perform an easy task (like distinguishing digits 'zero' to 'nine'[1]) with a very mediocre 95% accuracy was rated better than one that achieved, say, 75% in a more challenging task like transcribing natural language dictation[1], let alone a stunningly impressive 25% in following threads in a cocktail party[1].

    Lesson: take all reports of how such systems perform with more than a pinch of salt. Ideas like "percent accurate" need more context than you'll ever get to become meaningful.

    [1] These tasks are not really representative of what I'm talking about, but to go into detail would be serious levels of TMI. I guess that's a variant of the same problem the journos face when reporting on facial recognition.

    1. Kane Silver badge

      Re: Metrics

      "...let alone a stunningly impressive 25% in following threads in a cocktail party"

      Did they make all the molecules in the hostess's undergarments leap simultaneously one foot to the left?

      Sorry, wrong sort of party, mine's the one with the restaurant bill in the pocket...

      1. Nick Kew Silver badge

        Re: Metrics

        Not molecules. Threads.

        Good cocktail party metric from Dorothy Parker:

        “I like to have a martini,

        Two at the very most.

        After three I'm under the table,

        after four I'm under my host.”

    2. Tom 7 Silver badge

      Re: Metrics

      I've done some playing with AI. One thing that struck me was that there seems to be no not-sure option in most of the open training sets. Most of the algorithms seem to give various confidence levels but none I've seen reject things. This means pictures of white noise with a touch of red in are identified as robins after training on birds, not because they look like robins but because that's the closest match and it has no option of 'haven't a fucking clue mate' so it chooses the highest probability match. If you give it that option it seems to help reducing false positives but I haven't the experience, the data or the computing power to rest further.

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Biting the hand that feeds IT © 1998–2019