Research map / Adversarial robustness and security / Attacks on text and graphs
Bigram and Unigram Based Text Attack via Adaptive Monotonic Heuristic Search
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use adaptive Monotonic Heuristic Search instead of a static word replacement order | Word-level text attacks · IMDB and AG's News | successful attack rate: RAND 82.42% vs WSA 77.98% (AG's News, Word CNN); U-MHS 85.76% vs PWWS 81.98% | Helped |
| Add HowNet sememe candidates to WordNet synonym candidates (ablation) | U-MHS · AG's News, Word CNN | successful attack rate: 85.76% -> 92.77% | Helped |
| Add HowNet sememe candidates to WordNet synonym candidates (ablation) | U-MHS · AG's News, Word CNN | word substitutions: 4.815 -> 4.16 | Helped |
| Add bigram substitutions to synonym-sememe unigram search (ablation) | HU-MHS · AG's News, Word CNN | successful attack rate: 92.77% -> 91.77% | Hurt |
| Add bigram substitutions to synonym-sememe unigram search (ablation) | HU-MHS · AG's News, Word CNN | word substitutions: 4.16 -> 4.32 | Hurt |
| Use unigram-only substitutions with WordNet candidates rather than the full hybrid method (ablation) | BU-MHS · AG's News, Word CNN | successful attack rate: 91.77% -> 85.76% | Hurt |
| Use unigram-only substitutions with WordNet candidates rather than the full hybrid method (ablation) | BU-MHS · AG's News, Word CNN | word substitutions: 4.32 -> 4.815 | Hurt |
| Use bigram substitutions instead of modifying phrase words independently | BU-MHS · Text examples on IMDB, AG's News, and Yahoo! Answers | semantic preservation | Helped |
| Use BU-MHS-generated examples rather than PWWS examples for adversarial retraining | Word CNN on AG's News · AG's News adversarial retraining | test accuracy after retraining | Helped |
| Attack with BU-MHS rather than PWWS after adversarial retraining | Retrained Word CNN · AG's News | successful attack rate: 77.36% -> 77.36% | Helped |
| Generate transferable adversarial examples with BU-MHS rather than the compared methods | Yahoo! Answers victim models · Yahoo! Answers transfer attack | victim model accuracy under attack | Helped |
| Use BU-MHS rather than PWWS for targeted attacks | AG's News victim models · AG's News, target label 0 | successful attack rate: 87.21% -> 94.67% (Word CNN); 32.61% -> 65.07% (Char CNN); 77.92% -> 87.87% (Word LSTM) | Helped |
Closest papers by meaning
- Semantic Representation Attack against Aligned Large Language Models (NeurIPS 2025 · similarity 0.87)
- Attack-in-the-Chain: Bootstrapping Large Language Models for Attacks Against Black-Box Neural Ranking Models (AAAI 2025 · similarity 0.86)
- RigorLLM: Resilient Guardrails for Large Language Models against Undesired Content (ICML 2024 · similarity 0.85)
- LimeAttack: Local Explainable Method for Textual Hard-Label Adversarial Attack (AAAI 2024 · similarity 0.85)
- MathAttack: Attacking Large Language Models towards Math Solving Ability (AAAI 2024 · similarity 0.85)
- Revisiting Character-level Adversarial Attacks for Language Models (ICML 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.