Does this rely on overfitting the models, or is this generally possible? Seems like the former.
Implementation of one-pixel adversarial attacks for image classification (uses differential evolution), on CIFAR10, with Keras.
https://github.com/Hyperparticle/one-pixel-attack-keras …pic.twitter.com/KQfAoexcqd
-
-
-
This doesn't rely on adapting the networks to the attack, and it is generally possible as long as you have free access to network inference. See https://arxiv.org/abs/1710.08864
End of conversation
New conversation -
-
-
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
I have a hard time believing dropout wouldn't solve this.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Multi resolution should help against adversial attack. Having multiple CNN designed &trained for different input resolution should at least give an improved class probability with adversial examples.
-
When rescaled, the attack should give different results. Do you know if anyone has ever tried this?
End of conversation
New conversation -
-
-
I love it! how can someone be like u?
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
AI Winter is Coming...
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.