Reproduce “Attacking interpretable NLP systems” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Eldor Abdukhamidov, Tamer Abuhmed, Joanna C. S. Santos, Mohammed Abuhamad

Researchers from Sungkyunkwan University, University of Notre Dame, and Loyola University developed AdvChar, a black-box adversarial attack that simultaneously misleads deep learning NLP classifiers and ensures the generated interpretation explanations remain highly similar to benign ones. The attack achieved up to 80% misclassification success rates with an average of 1.63 to 4.17 character changes per input, demonstrating its ability to exploit interpretability features for stealthier attacks.

Key insights

  • Interpretable NLP systems (INLPS) designed to foster trust can introduce new vulnerabilities, as their transparency can be exploited by adversaries.
  • Black-box adversarial attacks can simultaneously mislead deep learning NLP classifiers and maintain high similarity in their interpretation explanations, rendering them stealthier and harder to detect.
  • Character-level perturbations, guided by interpreter output and constrained by interpretation similarity metrics, are highly effective and efficient in launching dual-objective attacks against INLPS.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.