Reproduce “BITE: Textual Backdoor Attacks with Iterative Trigger Injection” 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

Jun Yan, Vansh Gupta, Xiang Ren

A new textual backdoor attack called BITE was developed by researchers at USC and IIT Delhi, leveraging iterative trigger word injection to achieve high attack success rates while maintaining stealthiness. It reached an attack success rate of 62.8% on SST-2, outperforming baselines, and a proposed defense, DeBITE, reduced these rates by identifying and removing words with strong label correlations.

Key insights

  • NLP models can be effectively backdoored by exploiting and enhancing 'spurious correlations' between a target label and a dynamically identified set of trigger words.
  • Iterative, contextualized word-level perturbations using masked language models can achieve a superior balance between attack effectiveness and stealthiness for textual backdoors.
  • Sophisticated backdoor attacks can maintain high success rates even with low poisoning ratios and in clean-label settings, presenting a practical and potent threat.

Summary by alphaXiv.

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