Loading...
- Type of Document: M.Sc. Thesis
- Language: Farsi
- Document No: 52621 (19)
- University: Sharif University of Technology
- Department: Computer Engineering
- Advisor(s): Soleymani Baghshah, Mahdieh
- Abstract:
- In recent years, neural networks have been widely used in most machine learning domains. However, it has been shown that these networks are vulnerable to adversarial examples. adversarial examples are small and imperceptible perturbations applied to the input which lead to producing wrong output and thus, fooling the network. This will become an important issue in security related applications of deep neural networks, such as self-driving cars and medical diagnostics. Since, in the wort-case scenario, even human lives could be threatened. Although, many works have focused on crafting adversarial examples for image data, only a few studies have been done on textual data due to the existing challenges. The key challenge is the discrete nature of textual data which makes the optimization problems difficult to solve. In this project, we will study the robustness of neural networks against adversarial examples in the text domain . In addition to proposing a method for generating per sample adversarial examples, for the first time in this domain, we introduce a method for crafting universal adversarial examples which are input-independent. Our proposed methods are developed on an iterative algorithm based on gradient projection which finds a sequence of words to be added to the input sequence. Moreover, in order to make the crafted adversarial examples more meaningful, we leverage a language model in our methods. Our experiments show that text classifiers are quite vulnerable to our proposed methods and their accuracy may drop to the least possible value. At last, we perform adversarial training as a defense against our crafted examples which makes our models robust to them
- Keywords:
- Deep Neural Networks ; Adversarial Example ; Robustness ; Language Model ; Text Classification