4 ms·
when they're generated digitally, it's surprisingly hard to get good robustness against adversarial examples, and people have been working on it for a while. of
by currymj 7y ago
when they're generated digitally, it's surprisingly hard to get good robustness against adversarial examples, and people have been working on it for a while. of course you can train to stop any specific input, but then new ones can be created for the newly-trained network.
there are techniques that can prevent this, but you pay a real price in overall performance, which may or may not be worth it.
of course it's harder to make physical objects work, but if somebody did manage to make a method for physical adversarial attacks that worked reliably, I would guess you'd run into the same problems. it's not at all clear that the endgame is that we can easily stop these attacks.
- bidart 7y agoThe issue with adversarial examples is defending against a wide class of examples, not a few specific images. Clothes change much slower than software updates, and defending against a few adversarial examples is relatively easy, so this shouldn't be a problem.
- Retric 7y agoNN updates can take a long time to be trained, verified, and deployed. Especially if your talking a wide scale network of cameras. Some of these approaches are 7+ years old and still work.
- clort 7y agolong term storage overcomes the deficiencies in todays software As soon as they update the software, they run it over the stored video and investigate the differences. When they identify particular evasion techniques, they can flag up all instances and see what was done under their cover.
- egd 7y ago>Clothes change much slower than software updates, I change my clothes at least once a day and add to my wardrobe on a ~weekly to ~monthly basis. Having been party to large system upgrades at even a reasonably competent company, I'd be shocked to find out that any surveillance system anywhere is matching that velocity.
- AnthonyMouse 7y ago> The issue with adversarial examples is defending against a wide class of examples, not a few specific images. So who says it has to be a few specific images? They could auto-generate a different selection to choose from for each purchaser and then print to order. Every individual article of clothing could be unique.