Stories
Slash Boxes
Comments

SoylentNews is people

SoylentNews is powered by your submissions, so send in your scoop. Only 15 submissions in the queue.
posted by janrinok on Wednesday January 10 2018, @04:34PM   Printer-friendly
from the do-you-see-what-I-see? dept.

Image recognition technology may be sophisticated, but it is also easily duped. Researchers have fooled algorithms into confusing two skiers for a dog, a baseball for espresso, and a turtle for a rifle. But a new method of deceiving the machines is simple and far-reaching, involving just a humble sticker.

Google researchers developed a psychedelic sticker that, when placed in an unrelated image, tricks deep learning systems into classifying the image as a toaster. According to a recently submitted research paper about the attack, this adversarial patch is "scene-independent," meaning someone could deploy it "without prior knowledge of the lighting conditions, camera angle, type of classifier being attacked, or even the other items within the scene." It's also easily accessible, given it can be shared and printed from the internet.


Original Submission

 
This discussion has been archived. No new comments can be posted.
Display Options Threshold/Breakthrough Mark All as Read Mark All as Unread
The Fine Print: The following comments are owned by whoever posted them. We are not responsible for them in any way.
  • (Score: 3, Insightful) by requerdanos on Thursday January 11 2018, @12:49AM (2 children)

    by requerdanos (5997) Subscriber Badge on Thursday January 11 2018, @12:49AM (#620739) Journal

    I'd probably say, "it's a photo of a sticker and banana."

    Perhaps you would; I don't know you (or if I do, don't know it). But I would think the typical person, when asked to identify an image, would focus on things that occupy the largest portion of the image (and are therefore "the subject") and then, describe those things in an identifying way.

    We are looking at the images as digital photos or captures, just as the algorithms are, and it's not obvious that the circular trainwreck garbage fire area of the image is a sticker (much less a toaster), or whether it's a really strange poker chip on top of a photo, or whether it's a weird branding icon superimposed by latest-ecommerce-store.com, or what.

    But unless we just read TFA or work with such images frequently, that circular subimage is not going to be what we trigger on and say "Hey! I identify that part! This is a picture of a sticker! Oh, and there's also a banana," as you suggest.

    Rather, probably we'd focus on the part that we instantly identify: The yellow fruit.

    So, on Family Feud, the board would look something like this...

    One hundred people surveyed...
    Top two answers on the board...

    |--How would you identify this photo?---|
    | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
    | XXXXX [.99.] Banana XXXXXXXXXXXXXXXXX |
    | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
    | XXXXX [..1.] Sticker (And Banana) XXX |
    | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
    |---------------------------------------|

    I think that's kind of the point of the finding, that the sticker makes the bot classify the image as something different than would a person.

    - The sticker does not have to cover up the banana (or other subject) to get the image classified as a toaster.
    - The sticker does not have to cover up even a very large percentage of the image to get the image classified as a toaster.
    - The sticker does not have to convincingly depict a toaster to get the image classified as a toaster.

    I don't think that's really clickbait, but rather, a somewhat interesting finding.

    Permit me if you will to put on a moving picture with sound, in your imagination.

    Picture in your head someone showing someone else a rectangle of stiff paper imprinted with an image. In this case, let's say the image is of a man holding, displaying, a bizarrely-colored circle such as that in the banana-toaster image. Now, in your mind's ear, hear that person say to the other, "Hey, look, this is my dad."

    Even if the thought occurs to the other person that "That's not your dad; your dad is a person with a mass much greater than this square of paper and of a different shape and composition," they are not likely to say it. But it probably won't even *occur* to that other person to say "No, that's not your dad, that's a freaking toaster." (Your potential answer, "That's not your dad, that's primarily a sticker, and also some guy," is also on the not likely list.) But the bot? "Dude, that's not your dad, that's a toaster, 90+% probability" every single time.

    Starting Score:    1  point
    Moderation   +1  
       Insightful=1, Total=1
    Extra 'Insightful' Modifier   0  
    Karma-Bonus Modifier   +1  

    Total Score:   3  
  • (Score: 2) by fliptop on Thursday January 11 2018, @05:32AM (1 child)

    by fliptop (1666) on Thursday January 11 2018, @05:32AM (#620812) Journal

    But I would think the typical person, when asked to identify an image, would focus on things that occupy the largest portion of the image

    So you're saying it's really a picture of a tabletop w/ a sticker and a banana?

    --
    Our Constitution was made only for a moral and religious people. It is wholly inadequate to the government of any other.
    • (Score: 2) by requerdanos on Thursday January 11 2018, @11:32AM

      by requerdanos (5997) Subscriber Badge on Thursday January 11 2018, @11:32AM (#620893) Journal

      So you're saying it's really a picture of a tabletop w/ a sticker and a banana?

      Sigh. Yes, it does appear that that's what I was saying. Probably not, however, what I should have said.