Loading...

Application of Sparse Representations in Adversarial Machine Learning

Noshahri, Ehsan | 2022

82 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 55685 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Babaiezadeh, Massoud
  7. Abstract:
  8. Deep neural networks have been shown to perform very well in many classical machine learning tasks, including classification. However, it has been shown that these models are vulnerable to very small, and often imperceptible, adversarial perturbations of their input data, which makes it difficult to apply neural networks in security-critical areas. Finding the sparse solution of an underdetermined system of linear equations, which is the basis of sparse representation theory, is of significant importance in signal processing. Since finding such a solution requires minimizing the ℓ0 norm of a vector, which in turn requires using a combinatorial search, several methods for ℓ0 norm approximation have been proposed to make finding such a solution easier. In this thesis, we make use of one such method for performing ℓ0 norm adversarial attacks on image classifiers, and show that our proposed method can outperform similar methods such as SparseFool and JSMA, by creating sparser perturbations for the input images. Furthermore, based on the sparsity property of images in an appropriate dictionary, we present a method for reducing the effect of adversarial perturbations on images and show that, in the presence of adversarial perturbations, this method can increase the accuracy of deep learning models to a high extent
  9. Keywords:
  10. Adversarial Machine Learning ; Adversarial Attacks ; Sparse Representation ; Deep Learning ; Adversarial Defense

 Digital Object List

 Bookmark

...see more