3 ms·
Not really relevant to this. That deals with images, where you can perturb a huge number of pixels slightly to exploit weird edge cases. Even linear classifiers
by Houshalter 11y ago
Not really relevant to this. That deals with images, where you can perturb a huge number of pixels slightly to exploit weird edge cases. Even linear classifiers break on it, e.g. https://karpathy.github.io/2015/03/30/breaking-convnets/ https://karpathy.github.io/2015/03/30/breaking-convnets/
- empiricus 11y agoI'm pretty sure that if you move just a little the image, the convnet is no longer fooled. So this kind of exploit works only for completely stationary images, and it is very to overcome.