Generating Adversarial Input Sequences Based on RNN
Deep LearningNLPSecurity
This article is a summary of the 6th assignment for the Fall 2018 Computational Linguistics course, with the task of generating adversarial input sequences based on RNN.
First, let me give an overall introduction to the concept of adversarial example attacks. Some researchers have found that although deep neural networks achieve very high accuracy, applying small perturbations to the input can cause the model's prediction results to be completely wrong. For images, such perturbations are usually so subtle that they are imperceptible (humans cannot intuitively perceive the perturbation in the image), yet these adversarial input examples can successfully fool deep learning models. This type of attack that deceives neural network models can be broadly divided into two categories: untargeted attacks only need to make the predicted result of the image inconsistent with the original label, while targeted attacks need to misclassify the image as a specific class. It is worth noting that researchers have found that generated adversarial examples are still effective on other models, which is known as the transferability of adversarial examples.