Research map / Adversarial robustness and security / Attacks on text and graphs
Learning From Dictionary: Enhancing Robustness of Machine-Generated Text Detection in Zero-Shot Language via Adversarial Training
Yuanfan Li, Qi Zhou, Zexuan Xie
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 |
|---|---|---|---|
| add LAAL on top of adversarial training with RPL (ablation) | TASTE (mBERT) · M4GT zero-shot accuracy | avg accuracy (%): 70.9 -> 76.2 | Improved |
| add LAAL on top of adversarial training with RPL (ablation) | TASTE (mBERT) · M4GT under attacks | mean ASR (%): 20.8 -> 18.0 | Improved |
| add adversarial training (RPL) to mBERT (ablation) | mBERT · M4GT under attacks | mean ASR (%): 35.5 -> 20.8 | Improved |
| add LAAL to adversarial training, German (ablation) | TASTE · M4GT German | accuracy (%): 50.3 -> 69.5 | Improved |
| use gradient-based importance instead of attention-based (ablation) | TASTE · M4GT under attacks | mean ASR (%): 38.0 -> 18.0 | Improved |
| use gradient-based importance instead of random positions (ablation) | TASTE · M4GT under attacks | mean ASR (%): 30.0 -> 18.0 | Improved |
| use gradient-based importance instead of perturbation-based (ablation) | TASTE · M4GT under attacks | mean ASR (%): 27.6 -> 18.0 | Improved |
| use gradient-based importance instead of random positions (ablation) | TASTE · M4GT clean accuracy | mean accuracy: 0.756 -> 0.762 | No clear effect |
| increase attack strength ratio from 0.05 to 0.25 (ablation) | TASTE · M4GT | avg accuracy (%): 76.2 -> 73.1 | Hurt |
| increase attack strength ratio from 0.05 to 0.25 (ablation) | TASTE · M4GT under attacks | avg ASR (%): 18.0 -> 34.2 | Hurt |
| introduce 10% wrong dictionary entries (ablation) | TASTE · M4GT | cross-language mean accuracy (%): 76.2 -> 70.9 | Hurt |
| increase dictionary swap rate to 90% (ablation) | TASTE · M4GT | cross-language mean accuracy (%): 76.2 -> 68.0 | Hurt |
Closest papers by meaning
- DetectRL: Benchmarking LLM-Generated Text Detection in Real-World Scenarios (NeurIPS 2024 · similarity 0.88)
- Adversarial Paraphrasing: A Universal Attack for Humanizing AI-Generated Text (NeurIPS 2025 · similarity 0.88)
- Language-Driven Anchors for Zero-Shot Adversarial Robustness (CVPR 2024 · similarity 0.88)
- Few-Shot Adversarial Prompt Learning on Vision-Language Models (NeurIPS 2024 · similarity 0.86)
- DALD: Improving Logits-based Detector without Logits from Black-box LLMs (NeurIPS 2024 · similarity 0.86)
- Fast-DetectGPT: Efficient Zero-Shot Detection of Machine-Generated Text via Conditional Probability Curvature (ICLR 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.
