arxivst stuff from arxiv that you should probably bookmark

Adequacy of the Gradient-Descent Method for Classifier Evasion Attacks

Abstract · Apr 6, 2017 04:35 ·

cs-cr cs-lg stat-ml

Arxiv Abstract

  • Yi Han
  • Benjamin I. P. Rubinstein

Despite the wide use of machine learning in adversarial settings including computer security, recent studies have demonstrated vulnerabilities to evasion attacks—carefully crafted adversarial samples that closely resemble legitimate instances, but cause misclassification. In this paper, (1) we analyse the effectiveness of the gradient-descent method—the leading approach for generating adversarial samples—against non-linear support vector machines, and conclude that carefully reduced kernel smoothness can significantly increase robustness to the attack; (2) we propose a quantity similar to margin that can efficiently predict potential susceptibility to gradient-descent attack, before the attack is launched; and (3) we design a new adversarial sample construction algorithm based on optimising the multiplicative ratio of class decision functions. Our results demonstrate that the new method not only increases the attack’s success rate, but also achieves success with less perturbation.

Read the paper (pdf) »