Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note2 min read

Vulnerability of Natural Language Classifiers to Evolutionary Generated Adversarial Text

Source

arXiv

arxiv.org/abs/2606.27215v1

Read the source

Source

Plain-Language Brief

Researchers have developed a new method called GAversary, which uses a Genetic Algorithm (GA) and GloVe embeddings to generate adversarial attacks on natural language models. These attacks can significantly reduce the accuracy of the target model when tested, but they also perturb more words and take slightly longer than other methods like BAE and A2T.

What the Document Says

  • Deep learning models are vulnerable to adversarial inputs, especially in NLP.
  • Adversarial attacks often involve small, semantically similar token replacements to fool NLP models.
  • Recent methods have become more precise by targeting specific vulnerable words.
  • The paper proposes GAversary, a hybrid Genetic Algorithm to generate adversarial attacks on natural language models.
  • GAversary can treat the target model as a black box and requires only the logit value output by the model.
  • GAversary uses GloVe embeddings for word replacements (mutation operator) to improve semantic similarity of adversarial examples.
  • GAversary is applied to several benchmark data sets and target models.
  • GAversary reduces the target model's accuracy on test data compared to BAE and A2T attacks.
  • GAversary perturbs just under twice as many words as other methods.
  • GAversary has a slightly lower semantic similarity to the original text.
  • GAversary takes around a 5% increase in run-time.

Key Concepts Mentioned

  • Adversarial attacks
  • Natural language models
  • Genetic Algorithm (GA)
  • GloVe embeddings
  • Benchmark data sets

Claims Made by the Document

  • GAversary can significantly reduce the accuracy of target models.
  • GAversary perturbs more words than other methods.

Evidence Used

  • GAversary reduced a 76.8% accuracy to 5.8% compared to BAE's 27.6%
  • GAversary perturbs just under twice as many words as the other methods

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

Deep learning models have achieved impressive performance across various fields but remain vulnerable to adversarial inputs, particularly in NLP, where such attacks can have significant real-world consequences. Adversarial attacks often involve small, semantically similar token replacements to fool NLP models, and recent methods have become more precise by targeting specific vulnerable words, often by exploiting some level of access to the model's internal structure. This paper proposes GAversary, a hybrid Genetic Algorithm (GA) to generate adversarial attacks on natural language models. The GA is able to treat the target model as a black box, requiring only the logit value output by the model to guide the search. GAversary differs from GAs previously proposed for this problem by using GloVe embeddings to propose word replacements (the mutation operator) to improve the semantic similarity of the adversarial examples. GAversary is applied to several benchmark data sets and well-known target models. GAversary is able to substantially reduce the target model's accuracy on test data compared to the BAE and A2T attacks compared against (in the best case, reducing a 76.8% accuracy to 5.8%, compared to BAE's 27.6%). The trade-off is that GAversary perturbs just under twice as many words as the other two methods, with a slightly lower semantic similarity to the original text and around a 5% increase in run-time.