NaturalAdversaries: Can Naturalistic Adversaries Be as Effective as Artificial Adversaries?
Preprint
Abstract
While a substantial body of prior work has explored adversarial example generation for natural language understanding tasks, these examples are often unrealistic and diverge from the real-world data distributions. In this work, we introduce a two-stage adversarial example generation framework (NaturalAdversaries), for designing adversaries that are effective at fooling a given classifier and demonstrate natural-looking failure cases that could plausibly occur during in-the-wild deployment of the models. At the first stage a token attribution method is used to summarize a given classifier's behaviour as a function of the key tokens in the input. In the second stage a generative model is conditioned on the key tokens from the first stage. NaturalAdversaries is adaptable to both black-box and white-box adversarial attacks based on the level of access to the model parameters. Our results indicate these adversaries generalize across domains, and offer insights for future research on improving robustness of neural text classification models.Keywords
All Related Versions
This publication has 23 references indexed in Scilit:
- Data Governance in the Age of Large-Scale Data-Driven Language Technology2022
- Connecting Attributions and QA Model Behavior on Realistic Counterfactuals2021
- Challenges in Automated Debiasing for Toxic Language Detection2021
- Learning from the Worst: Dynamically Generated Datasets to Improve Online Hate Detection2021
- How Furiously Can Colorless Green Ideas Sleep? Sentence Acceptability in ContextTransactions of the Association for Computational Linguistics, 2020
- Is BERT Really Robust? A Strong Baseline for Natural Language Attack on Text Classification and EntailmentProceedings of the AAAI Conference on Artificial Intelligence, 2020
- Social Bias Frames: Reasoning about Social and Power Implications of Language2020
- Linguistic generalization and compositionality in modern artificial neural networksPhilosophical Transactions of the Royal Society B: Biological Sciences, 2019
- Still a Pain in the Neck: Evaluating Text Representations on Lexical CompositionTransactions of the Association for Computational Linguistics, 2019
- Data Statements for Natural Language Processing: Toward Mitigating System Bias and Enabling Better ScienceTransactions of the Association for Computational Linguistics, 2018