Universal Adversarial Triggers for Attacking and Analyzing NLP

TL;DR

This paper introduces universal adversarial triggers that can manipulate NLP models across tasks, revealing vulnerabilities and biases, and providing insights into model behavior through input-agnostic triggers.

Contribution

The paper proposes a gradient-guided method to find universal triggers that attack NLP models and analyze their behavior, demonstrating transferability and interpretability.

Findings

Triggers drastically reduce model accuracy on targeted tasks.

Triggers transfer across different models and tasks.

Triggers reveal dataset biases and model heuristics.

Abstract

Adversarial examples highlight model vulnerabilities and are useful for evaluation and interpretation. We define universal adversarial triggers: input-agnostic sequences of tokens that trigger a model to produce a specific prediction when concatenated to any input from a dataset. We propose a gradient-guided search over tokens which finds short trigger sequences (e.g., one word for classification and four words for language modeling) that successfully trigger the target prediction. For example, triggers cause SNLI entailment accuracy to drop from 89.94% to 0.55%, 72% of "why" questions in SQuAD to be answered "to kill american people", and the GPT-2 language model to spew racist output even when conditioned on non-racial contexts. Furthermore, although the triggers are optimized using white-box access to a specific model, they transfer to other models for all tasks we consider. Finally, since triggers are input-agnostic, they provide an analysis of global model behavior. For instance, they confirm that SNLI models exploit dataset biases and help to diagnose heuristics learned by reading comprehension models.