Research map / Adversarial robustness and security / Attacks on text and graphs
TextShield: Beyond Successfully Detecting Adversarial Sentences in text classification
Lingfeng Shen, Ze Zhang, Haiyun Jiang, Ying Chen
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 TextShield instead of SEM | BERT defense · IMDB adversarial examples (IGA, PWWS, GA, TextFooler) | accuracy: 87.5 -> 88.4; 88.4 -> 88.6; 87.3 -> 90.5; 85.5 -> 86.5 | Improved |
| Use TextShield instead of SEM | BERT defense · AGNews adversarial examples (IGA, PWWS, GA, TextFooler) | accuracy: 88.5 -> 90.2; 88.5 -> 90.0; 88.5 -> 90.0; 88.5 -> 89.5 | Improved |
| Use TextShield instead of SEM | BERT defense · Yahoo adversarial examples (IGA, PWWS, GA, TextFooler) | accuracy: 66.8 -> 70.5; 66.8 -> 69.0; 66.4 -> 70.0; 62.0 -> 64.5 | Improved |
| Remove the VG sub-detector (ablation) | TextShield · AGNews, TextCNN, TL/PWWS/GA/IGA | accuracy: -8.9; -7.9; -6.4; -7.0 | Hurt |
| Add the VG sub-detector (ablation) | TextShield · AGNews, TextCNN, TL/PWWS/GA/IGA | accuracy: +8.9; +7.9; +6.4; +7.0 | Improved |
| Remove all sub-detectors (ablation) | TextShield · AGNews, TextCNN, TL/PWWS/GA/IGA | accuracy: -34.8; -29.6; -40.3; -36.9 | Hurt |
| Add all four saliency sub-detectors (ablation) | TextShield · AGNews, TextCNN, TL/PWWS/GA/IGA | accuracy: +34.8; +29.6; +40.3; +36.9 | Improved |
| Set corrector threshold β to 0.4 instead of a small or large value (ablation) | TextShield corrector · AGNews, three victim classifiers | accuracy | Improved |
| Increase the number of adversarial training sentences per class K (ablation) | TextShield detector · TextShield defense configurations | accuracy | Improved |
| Increase K beyond 2000 adversarial sentences per class (ablation) | TextShield detector · TextShield defense configurations | accuracy | No clear effect |
Later papers that compared with this method
1 later accepted paper reports TextShield as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Guided Perturbation Sensitivity (GPS): Detecting Adversarial Text via Embedding Stability and Word Importance (AAAI 2026 · better on 5 of 6 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Adversarial Paraphrasing: A Universal Attack for Humanizing AI-Generated Text (NeurIPS 2025 · similarity 0.88)
- DetectRL: Benchmarking LLM-Generated Text Detection in Real-World Scenarios (NeurIPS 2024 · similarity 0.86)
- ToxicTextCLIP: Text-Based Poisoning and Backdoor Attacks on CLIP Pre-training (NeurIPS 2025 · similarity 0.85)
- OUTFOX: LLM-Generated Essay Detection Through In-Context Learning with Adversarially Generated Examples (AAAI 2024 · similarity 0.85)
- GuardT2I: Defending Text-to-Image Models from Adversarial Prompts (NeurIPS 2024 · similarity 0.85)
- Revisiting Character-level Adversarial Attacks for Language Models (ICML 2024 · similarity 0.85)
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.
