0.25
0.5
0.75
1.25
1.5
1.75
2
Adversarial Examples
Published on Sep 13, 201510911 Views
Related categories
Chapter list
Adversarial Examples00:00
In this presentation….00:17
Universal engineering machine01:16
Deep neural networks are as good as humans at...03:12
Do neural networks “understand” these tasks?04:11
Turning objects into “airplanes”05:45
Attacking a linear model09:56
Clever Hans11:38
Adversarial examples from overfitting14:12
Adversarial examples from underfitting16:23
Different kinds of low capacity17:38
Modern deep nets are very (piecewise) linear19:51
Not every class change is a mistake26:09
Guaranteeing that class changes are mistakes27:34
A thin manifold of accuracy34:39
The Fast Gradient Sign Method35:18
High-dimensional linear models35:54
Higher-dimensional linear models35:55
Linear Adversarial examples37:53
RBFs behave more intuitively far from the data38:31
Easy to optimize = easy to perturb38:45
Ubiquitous hallucinations42:52
Methods based on expensive search, strong hand-designed priors42:56
Cross-model, cross-dataset generalization - 150:53
Cross-model, cross-dataset generalization - 252:40
Adversarial examples in the human visual system57:19
Failed defenses - 201:07:59
Security implications01:08:40
Universal approximator theorem01:08:41
Training on adversarial examples01:09:36
Generative modeling cannot solve the problem01:13:42
Weaknesses persist01:19:57
Pertubation’s effect on class distributions01:20:53
Pertubation’s effect after adversarial training01:22:51
Virtual adversarial training01:23:29
Failed defenses - 101:25:16
Please use evidence, not speculation01:26:57
Recommended adversarial example benchmark01:26:59
Alternative adversarial example benchmark01:28:38
Recommended fooling image / rubbish class benchmark01:29:00