Research map / Adversarial robustness and security / Attacks on text and graphs
Punctuation-level Attack: Single-shot and Single Punctuation Can Fool Text Models
Wenqiang Wang, Chongyang Du, Tao Wang, Kaihao Zhang, Wenhan Luo, Lin Ma, Wei Liu, Xiaochun Cao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 Insertion instead of Deletion attack mode (ablation) | punctuation-level attack · CoLA, ELECTRA | fool rate (Top-1): 5.18% -> 67.40% | Improved |
| use Insertion traversal search instead of single-shot Top-1 (ablation) | punctuation insertion attack · CoLA, ELECTRA | fool rate: 67.40% -> 90.80% | Improved |
| increase Top-k candidates from Top-1 to Top-3 and Top-5 (ablation) | punctuation insertion attack · CoLA, ELECTRA | fool rate: 67.40% -> 73.06% -> 73.83% | Improved |
| increase average search space (CoLA 362.62 to Wanli 1161.12) (ablation) | single-shot Insertion attack · Wanli, RoBERTa | fool rate (Top-1): 67.40% -> 8.44% | Hurt |
| use Displacement instead of Insertion for single shot (ablation) | punctuation attack · CoLA, XLMR | fool rate (Top-1): 28.76% -> 43.05% | Improved |
| use Replacement instead of Insertion for single shot (ablation) | punctuation attack · Wanli, RoBERTa | fool rate (Top-1): 8.44% -> 8.48% | No clear effect |
| apply Insertion attack Top-1 (ablation) | Sentence-BERT · STS-B | Pearson correlation: 0.7990 -> 0.7874 | Hurt |
| apply Insertion attack traversal (ablation) | Sentence-BERT · STS-B | Pearson correlation: 0.7990 -> 0.6992 | Hurt |
| apply Insertion attack Top-1 (ablation) | Distilbert · STS-B | Spearman correlation: 0.7257 -> 0.7035 | Hurt |
| apply Insertion attack Top-1 (ablation) | summarization model · Gigaword | ROUGE-1: 11.69 -> 10.91 | Hurt |
| apply Insertion attack traversal (ablation) | summarization model · Gigaword | ROUGE-1: 11.69 -> 5.22 | Hurt |
| insert a punctuation into prompt (Top-1) (ablation) | Stable Diffusion V2 · text-to-image | CLIP score: 0.3278 -> 0.3176 | Hurt |
Closest papers by meaning
- Adversarial Paraphrasing: A Universal Attack for Humanizing AI-Generated Text (NeurIPS 2025 · similarity 0.88)
- ToxicTextCLIP: Text-Based Poisoning and Backdoor Attacks on CLIP Pre-training (NeurIPS 2025 · similarity 0.86)
- LimeAttack: Local Explainable Method for Textual Hard-Label Adversarial Attack (AAAI 2024 · similarity 0.84)
- Semantic Representation Attack against Aligned Large Language Models (NeurIPS 2025 · similarity 0.84)
- Revisiting Character-level Adversarial Attacks for Language Models (ICML 2024 · similarity 0.84)
- An LLM can Fool Itself: A Prompt-Based Adversarial Attack (ICLR 2024 · similarity 0.83)
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.
