fast gradient sign method

Given input $x$ with target $y$ , and loss function $\mathcal{L}$ , an adversarial example can be defined as $x\prime = x + \varepsilon \mathrm{sign}(\nabla_x \mathcal{L}$ .

Backlinks

deep-learning-model-abstraction

- Both conventional classifier and [[fast gradient sign method]] like methods.

ML Reviews

- [[fast gradient sign method]]

scalable-uncertainties-from-deep-ensembles

Basically use conventional [[fast gradient sign method]], or [[virtual adversarial training]].

Adversarial training

Goodfellow proposed the [[fast gradient sign method]] to generate adversarial examples quickly. Under that framework, with $\varepsilon$ noise, you basically smooth out your regressor predictions into a $\varepsilon$-neighborhood.