GLARE: Generative Left-to-right AdversaRial Examples.

Recently, transformer models [1] have been applied to adversarial example generation—word level substitution models utilizing BERT [2] ((3], [4], [5]) have out-performed previous state-of-the-art approaches. Extending the paradigm of transformer-based generation of adversarial examples, we propose a novel textual adversarial example generation framework based on transformer language models: our method (GLARE) generates wordand span-level perturbations of input examples using ILM [6], a GPT-2 language model finetuned to fill in masked spans. We demonstrate that GLARE achieves a superior performance to CLARE (the current state-of-the-art model) in terms of attack success rate and semantic similarity between the perturbed and original examples. 1 Key Information to include ¢ CS 224N Mentor: Shikhar Murty e Stanford AI Mentor: Ethan A. Chi e External Collaborators: N/A e Sharing Project: N/A