First-Order Adversarial Vulnerability of Neural Networks and Input Dimension

110 citations recorded by the source corpus.

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

Adversarial Defense Via Local Flatness Regularization (2019, unknown): “Specifically, in [13], the relationship between the average value of the chosen norm of the gradient regarding to input space and the adversarial vulnerability was discussed, based on which the authors proposed a defense method by regularizing that value.”
Towards Deep Learning Models Resistant to Large Perturbations (2020, background): “The effect of different initializations on adversarial training has not been studied rigorously in the literature (Ben Daya et al., 2018; Simon-Gabriel et al., 2018).”
Boosting Adversarial Training with Hypersphere Embedding (2020, background): “Most of the adversarial attacks applied in AT belong to the family of first-order adversaries [65], due to the computational efficiency.”
A Kernel Perspective for Regularizing Deep Neural Networks (2018, unknown): “Related penalties have been considered in the context of robust optimization, for regularization or robustness, noting that a penalty based on the gradient of the loss function x 7→ l(y, f(x)) can give a good approximation of (7) when ǫ is small (Drucker & Le Cun, 1991; Lyu et al., 2015; Roth et al.”
Understanding and Improving Fast Adversarial Training (2020, background): “(1) and instead regularize the model to improve robustness [36, 25, 28], however this does not lead to higher robustness compared to standard adversarial training.”