Research map / Adversarial robustness and security / Attacks on text and graphs
Guided Perturbation Sensitivity (GPS): Detecting Adversarial Text via Embedding Stability and Word Importance
Bryan E. Tuck, Rakesh M. Verma
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 gradient attribution instead of random word ranking (ablation) | GPS · Adversarial detection (Table 3) | detection accuracy: 0.694 -> 0.836 (Yelp RoBERTa BERT-Attack) | Improved |
| use gradient attribution instead of attention rollout (ablation) | GPS · Adversarial detection (word-level attacks) | detection accuracy: 0.714 -> 0.836 (Yelp RoBERTa BERT-Attack) | Improved |
| use attention rollout instead of gradient on char-level attack (ablation) | GPS · IMDB DeepWordBug | detection accuracy: 0.935 vs 0.958 (RoBERTa) | No clear effect |
| guided importance masking vs random masking: sensitivity ratio (ablation) | GPS features · Embedding sensitivity (Table 1) | adv/benign ratio: 1.932 vs 1.880 | No clear effect |
| GPS (Gradient) vs TextShield | GPS · IMDB DeBERTa BERT-Attack | detection accuracy: 0.731 -> 0.755 | Improved |
| GPS (Gradient) vs Sharpness baseline | GPS · IMDB DeBERTa BERT-Attack | detection accuracy: 0.797 vs 0.755 | Hurt |
| GPS (Gradient) vs Sharpness baseline | GPS · Yelp RoBERTa BERT-Attack | detection accuracy: 0.910 vs 0.836 | Hurt |
| transfer detector from DeepWordBug to BERT-Attack | GPS (Gradient) · Cross-attack transfer | F1: 0.943 -> 0.649 | Hurt |
| reduce masked words K from 50 to 5 (ablation) | GPS · Detection F1 | F1: <0.015 F1 variation | No clear effect |
| train on DeBERTa test on RoBERTa | GPS (Gradient) · Cross-model transfer | F1: 0.852 -> 0.880 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- TextShield — TextShield: Beyond Successfully Detecting Adversarial Sentences in text classification (ICLR 2023) · 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
- RobustStress: Stress-Testing AI-Generated Text Detectors under Gradual Perturbations (NeurIPS 2026 · similarity 0.89)
- Detecting Adversarial Data Using Perturbation Forgery (CVPR 2025 · similarity 0.88)
- Word Level Robustness Enhancement: Fight Perturbation with Perturbation (AAAI 2022 · similarity 0.87)
- Adversarial Paraphrasing: A Universal Attack for Humanizing AI-Generated Text (NeurIPS 2025 · similarity 0.87)
- TextGrad: Advancing Robustness Evaluation in NLP by Gradient-Driven Optimization (ICLR 2023 · similarity 0.87)
- Perturbation-Invariant Adversarial Training for Neural Ranking Models: Improving the Effectiveness-Robustness Trade-Off (AAAI 2024 · similarity 0.86)
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.
