4 ms·

See e.g. https://towardsdatascience.com/breaking-neural-networks-with-adversarial-attacks-f4290a9a45aa https://towardsdatascience.com/breaking-neural-networks-w

• jasode 8y ago
  I wasn't the one who downvoted you but I agree with IanCal and this blog title didn't trigger my "clickbait" sensor. I disagree that this is an example of "adversarial attack". The famous duck/rabbit illusion has been around since ~1892[1] and therefore was not deliberately constructed to be an "adversary" to image classification neural networks. To me, it's an interesting example of feeding a well-known optical illusion to an AI algorithm and observing its behavior. [1] https://en.wikipedia.org/wiki/Rabbit%E2%80%93duck_illusion https://en.wikipedia.org/wiki/Rabbit%E2%80%93duck_illusion
  • stared 8y ago
    Yes, I agree with you. (No idea who downvoted my and why either.) This (well known) illusion is NOT an adversarial example. Though, I explain why for people working with AI (e.g. me) the title seemed like mentioning an adversarial example. There are plenty of examples of "just rotate and a vulture becomes an orangutan" where it does not look like an orangutan for humans. Vide: "A Rotation and a Translation Suffice: Fooling CNNs with Simple Transformations" https://arxiv.org/pdf/1712.02779.pdf https://arxiv.org/pdf/1712.02779.pdf