Universal Adversarial Attack Using Very Few Test ExamplesDownload PDF

25 Sep 2019 (modified: 24 Dec 2019)ICLR 2020 Conference Blind SubmissionReaders: Everyone
  • Original Pdf: pdf
  • Keywords: universal, adversarial, SVD
  • Abstract: Adversarial attacks such as Gradient-based attacks, Fast Gradient Sign Method (FGSM) by Goodfellow et al.(2015) and DeepFool by Moosavi-Dezfooli et al. (2016) are input-dependent, small pixel-wise perturbations of images which fool state of the art neural networks into misclassifying images but are unlikely to fool any human. On the other hand a universal adversarial attack is an input-agnostic perturbation. The same perturbation is applied to all inputs and yet the neural network is fooled on a large fraction of the inputs. In this paper, we show that multiple known input-dependent pixel-wise perturbations share a common spectral property. Using this spectral property, we show that the top singular vector of input-dependent adversarial attack directions can be used as a very simple universal adversarial attack on neural networks. We evaluate the error rates and fooling rates of three universal attacks, SVD-Gradient, SVD-DeepFool and SVD-FGSM, on state of the art neural networks. We show that these universal attack vectors can be computed using a small sample of test inputs. We establish our results both theoretically and empirically. On VGG19 and VGG16, the fooling rate of SVD-DeepFool and SVD-Gradient perturbations constructed from observing less than 0.2% of the validation set of ImageNet is as good as the universal attack of Moosavi-Dezfooli et al. (2017a). To prove our theoretical results, we use matrix concentration inequalities and spectral perturbation bounds. For completeness, we also discuss another recent approach to universal adversarial perturbations based on (p, q)-singular vectors, proposed independently by Khrulkov & Oseledets (2018), and point out the simplicity and efficiency of our universal attack as the key difference.
4 Replies