Post ASr8UeNKrmabe5nC7M by wasootch@mastodon.sdf.org
(DIR) More posts by wasootch@mastodon.sdf.org
(DIR) Post #ASqdiOqeIP7Kxpa0zA by clive@saturation.social
2023-02-19T20:10:37Z
0 likes, 2 repeats
Behold "Glaze", a software tool that subtly alters the pixels in a piece of art online 🎨 ...... in a way that isn't detectable by the human eye 👀 ...... but renders the image useless for training a visual AI. The AI can't detect the artists' style, and thus can't mimic it and copy it. 🤖 ✖️ It's item #1 in my weekly "Linkfest" newsletter here 📫 : https://buttondown.email/clivethompson/archive/linkfest-6-gps-on-the-moon-8-bit-jane-austen-and/(Subscribe here! https://buttondown.email/clivethompson)
(DIR) Post #ASqt2bfdpYNzpuWaVE by air_adair@mas.to
2023-02-19T23:02:07Z
0 likes, 0 repeats
@clive is it strange that i sort of like some of the “ruined” renders? but at least they aren’t infringing on the artists!
(DIR) Post #ASqvIIRFIq5MjLOt0q by shyduroff@mastodon.mit.edu
2023-02-19T23:27:33Z
0 likes, 0 repeats
@clive … interesting …
(DIR) Post #ASqzjoLBCP3hJBwF8K by Badscrew@hachyderm.io
2023-02-20T00:17:32Z
0 likes, 0 repeats
@clive " isn't detectable by the human eye" I can clearly see the difference?
(DIR) Post #ASr8UeNKrmabe5nC7M by wasootch@mastodon.sdf.org
2023-02-20T01:55:35Z
0 likes, 0 repeats
@clive It's a bit late now though. But this is good I guess.
(DIR) Post #ASrKJKsWDMrpRPkWQq by paulawirth@tribe.net
2023-02-20T04:08:00Z
0 likes, 0 repeats
@clive I am very excited about this kind of technology. Artists need be able to protect their work (and style) from AI scraping.
(DIR) Post #ASrLKyN4xyMYAbO8Gm by itscooldani@mstdn.social
2023-02-20T04:19:33Z
0 likes, 0 repeats
@clive I would love a before/after of the same artist if you have one? It’s hard for me to tell the effect from separate ones.
(DIR) Post #ASrYU3U8ca1Q2aLpmy by eob@social.coop
2023-02-20T06:46:48Z
0 likes, 0 repeats
@clive I'm a bit skeptical that the Glaze approach will work long term as the visual AI improvesIt only works now because the AI takes into account features that are invisible to humans when it trains on the visual artBut that presumably is a flaw in the AI, given it's targeting human perceptionA better AI would only take into account features that humans perceiveAnd then the trick of making imperceptible changes to fool the AI would not work, as those changes would be ignored by the AI
(DIR) Post #ASrugJQDOPHvvylYAa by Wyndix@fosstodon.org
2023-02-20T10:55:35Z
0 likes, 0 repeats
@clive seems like it's easy to bypass though :blobcatThink:
(DIR) Post #ASrunMyMabGGyR6UZE by legumancer@tech.lgbt
2023-02-20T10:56:20Z
0 likes, 0 repeats
@clive They did it! Somebody else built the thing I'd been wishing I could build!
(DIR) Post #ASs4H5xvHjKEVC9ANk by lilianedwards@someone.elses.computer
2023-02-20T12:43:02Z
0 likes, 0 repeats
@clive wow. I wondr how far the years of anti cctv surveillance work is feeding into the new anti-datamining resistance...
(DIR) Post #ASs89qS8xgGcBF53jc by tooterphone_tooterflute@mstdn.social
2023-02-20T13:26:33Z
0 likes, 0 repeats
@cliveThanks for sharing!#AI #Glaze #AIArt
(DIR) Post #ASsEIJ0RHjHEmIXW1g by smallfreak@social.vivaldi.net
2023-02-20T14:34:06Z
0 likes, 0 repeats
@clive well, "can't" is certainly a factor of time and evolvement. There are a zillion things an AI "can't do" - if you read statements from 5 years ago. And sure they do by now.There is NOTHING that a human can do that an AI can't by principle. You can safely say "not yet". Maybe some time it's quite the opposite. Other than the human brain, an AI can be scaled without technical limitation.
(DIR) Post #ASspvf098aEkap21WS by TrashPanda@infosec.exchange
2023-02-20T21:36:58Z
0 likes, 0 repeats
@clive I dunno, only the bluhm one looked like it was really failing. Maybe I just don't have an eye for art, or I'm the only one that thinks if GLAZE works now they'll be able to bypass it in a couple more versions of they want. Plus they already have access to the unaltered versions of all currently famous artists' work...
(DIR) Post #ASsyBG6wFLXpfB9pJI by vespasianvs@mastodon.online
2023-02-20T23:08:26Z
0 likes, 0 repeats
@clive that is pretty cool! However, where can I get a copy of the one on the bottom right? I'll have to confess I love that one 😂
(DIR) Post #ASwcU4izNrHTUZMWOm by clive@saturation.social
2023-02-22T17:24:07Z
0 likes, 0 repeats
@vespasianvs Agreed, they have a real charm
(DIR) Post #ASwcrUqIRbsZ9hwrNQ by clive@saturation.social
2023-02-22T17:24:34Z
0 likes, 0 repeats
@legumancer Aha, cool, you had the idea too!
(DIR) Post #ASwd0nxgjFdR3iqPVQ by clive@saturation.social
2023-02-22T17:24:23Z
0 likes, 0 repeats
@lilianedwards I wonder too!
(DIR) Post #ASwhYr1YFWC8De8upc by legumancer@tech.lgbt
2023-02-22T18:22:03Z
0 likes, 0 repeats
@clive building it is the important part, though!
(DIR) Post #ASySS8GcdDZCfrRs9Y by BillySmith@social.coop
2023-02-20T07:06:18Z
0 likes, 0 repeats
@eob @clive It's just the beginning of another arms race. :D
(DIR) Post #ASySS95fZQOrEBKdii by BillySmith@social.coop
2023-02-20T07:44:16Z
0 likes, 0 repeats
@eob @clive This is also similar to the printing protocols used by the high-quality photocopiers where there is a hidden watermark on banknotes that when scanned by a photocopier, makes the photocopier refuse to copy it. :D
(DIR) Post #ASySS9gBNfbh3QPndI by bornach@fosstodon.org
2023-02-20T08:53:13Z
0 likes, 0 repeats
@BillySmith @eob @cliveThat requires the cooperation of the printer/scanner manufacturershttps://murdoch.is/projects/currency/Glaze seems to be using an approach similar to adversarial images designed to make deep neural networks trained for object recognition tasks to misidentify what they seehttps://towardsdatascience.com/breaking-neural-networks-with-adversarial-attacks-f4290a9a45aa
(DIR) Post #ASySSA8BhY0aSHWAHg by nicemicro@fosstodon.org
2023-02-23T09:32:31Z
0 likes, 0 repeats
@bornach @BillySmith @eob @clive I imagine these alterations have to be specific to the models it tries to protect against? or am I mistaken here?I specifically remember that for the object recognition engines could be improved by adding the manipulated images to the training set.
(DIR) Post #ASySSAenkI625QmD7Q by BillySmith@social.coop
2023-02-23T10:41:58Z
0 likes, 0 repeats
@nicemicro @bornach @eob @clive That sort of feedback mechanism is always susceptible to a Godel-ian attack. :DAlways approaching perfection, but never getting there. :D
(DIR) Post #ASySSB7s0DLfXaNQQa by BillySmith@social.coop
2023-02-23T10:44:02Z
0 likes, 1 repeats
@nicemicro @bornach @eob @clive Just saw this in my feed:https://social.coop/@muffinista@toot.site/109910350502348235 :D