Detection of Adversarial Examples in NLP: Benchmark and Baseline via Robust Density Estimation

Anonymous

Abstract: Word-level adversarial attacks have shown success in NLP models, drastically decreasing the performance of transformer-based models in recent years. As a counter measure, adversarial defense has been explored, but relatively little efforts have been made to detect adversarial examples. However, detecting adversarial examples in NLP may be crucial for automated task (e.g. review sentiment analysis) that wishes to amass information about a certain population and additionally be a step towards a robust defense system. To this end, we release a dataset for four popular attack methods on three datasets and four NLP models to encourage further research in this field. Along with it, we propose a competitive baseline based on density estimation that has the highest \textsc{auc} on 21 out of 22 dataset-attack-model combinations.\footnote{https://github.com/anoymous92874838/text-adv-detection}