eduzhai > Applied Sciences > Engineering >

Adversarially Training for Audio Classifiers

  • king
  • (0) Download
  • 20210507
  • Save

... pages left unread,continue reading

Document pages: 8 pages

Abstract: In this paper, we investigate the potential effect of the adversariallytraining on the robustness of six advanced deep neural networks against avariety of targeted and non-targeted adversarial attacks. We firstly show that,the ResNet-56 model trained on the 2D representation of the discrete wavelettransform appended with the tonnetz chromagram outperforms other models interms of recognition accuracy. Then we demonstrate the positive impact ofadversarially training on this model as well as other deep architecturesagainst six types of attack algorithms (white and black-box) with the cost ofthe reduced recognition accuracy and limited adversarial perturbation. We runour experiments on two benchmarking environmental sound datasets and show thatwithout any imposed limitations on the budget allocations for the adversary,the fooling rate of the adversarially trained models can exceed 90 . In otherwords, adversarial attacks exist in any scales, but they might require higheradversarial perturbations compared to non-adversarially trained models.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×