HotFlip: White-Box Adversarial Examples for Text Classification

• DOI: 10.18653/v1/p18-2006
• Publication Date: 2019-06-29T15:52:10Z
• Authors: Javid Ebrahimi; Anyi Rao; Daniel Lowd; Dejing Dou
• Keywords: White box

We propose an efficient method to generate white-box adversarial examples trick a character-level neural classifier. find that only few manipulations are needed greatly decrease the accuracy. Our relies on atomic flip operation, which swaps one token for another, based gradients of one-hot input vectors. Due efficiency our method, we can perform training makes model more robust attacks at test time. With use semantics-preserving constraints, demonstrate HotFlip be adapted attack word-level classifier as well.