3 ms·
If your definition of "hack" is "get bob to accept bad file", no, this model is not easier - it's just different. You could literally piggyback on the director
by Klonoar 5y ago
If your definition of "hack" is "get bob to accept bad file", no, this model is not easier - it's just different.
You could literally piggyback on the directories that Macs use to sync to iCloud Drive, get an image in there, and then it gets scanned by iCloud. This is not some new theoretical attack - and in fact, this would be the "hack" for the new one as well since it requires iCloud sync to trigger anyway.
- visarga 5y agoI was referring to a "white box" adversarial attack, meaning you can try variations of the input to see how the outcome changes, and then construct an input that will fool the system. That's only possible if you have access to the perceptual hashing model, which you do now since it runs locally. Of course generating an adversarial image is not the final step, the hacker still need to place it in the victim's account somehow, but it's easier now because the image looks legit.