Research map / Adversarial robustness and security / Backdoor and poisoning
ParaFuzz: An Interpretability-Driven Technique for Detecting Poisoned Samples in NLP
Lu Yan, ZHUO ZHANG, Guanhong Tao, Kaiyuan Zhang, Xuan Chen, Guangyu Shen, Xiangyu Zhang
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 ParaFuzz instead of ONION | detection of style backdoor · SST-2 style attack | F1 (%): 57.7 -> 89.6 | Improved |
| use ParaFuzz instead of ONION | detection of Hidden Killer attack · AGNews | F1 (%): 10.2 -> 77.6 | Improved |
| use ParaFuzz instead of ONION | detection of Embedding-Poisoning · IMDB | F1 (%): 94.2 -> 93.4 | No clear effect |
| use ParaFuzz instead of ONION | detection of Badnets, TrojAI model 12 · TrojAI round 6 | F1 (%): 81.1 -> 93.0 | Improved |
| use ParaFuzz instead of ONION | detection of Badnets, TrojAI model 41 · TrojAI round 6 | F1 (%): 75.1 -> 79.4 | Improved |
| use ParaFuzz instead of ONION | detection of Badnets, TrojAI model 46 · TrojAI round 6 | F1 (%): 91.9 -> 87.1 | Hurt |
| add keyword-based mutation (ablation) | ParaFuzz · TrojAI model 12 | F1 (%): 91.8 -> 93.0 | Improved |
| add keyword-based mutation (ablation) | ParaFuzz · TrojAI model 17 | F1 (%): 95.8 -> 92.9 | Hurt |
| add structure-based mutation (ablation) | ParaFuzz · TrojAI model 20 | F1 (%): 85.3 -> 92.9 | Improved |
| add structure-based mutation (ablation) | ParaFuzz · TrojAI model 12 | F1 (%): 90.6 -> 93.0 | Improved |
| add evolutionary mutation (ablation) | ParaFuzz · TrojAI model 12 | F1 (%): 87.7 -> 93.0 | Improved |
| add evolutionary mutation (ablation) | ParaFuzz · TrojAI model 15 | F1 (%): 92.7 -> 91.7 | Hurt |
Closest papers by meaning
- Probe before You Talk: Towards Black-box Defense against Backdoor Unalignment for Large Language Models (ICLR 2025 · similarity 0.88)
- BackdoorLLM: A Comprehensive Benchmark for Backdoor Attacks and Defenses on Large Language Models (NeurIPS 2025 · similarity 0.88)
- Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples (NeurIPS 2026 · similarity 0.88)
- Demystifying Poisoning Backdoor Attacks from a Statistical Perspective (ICLR 2024 · similarity 0.87)
- Simulate and Eliminate: Revoke Backdoors for Generative Large Language Models (AAAI 2025 · similarity 0.87)
- Defense against Backdoor Attack on Pre-trained Language Models via Head Pruning and Attention Normalization (ICML 2024 · similarity 0.87)
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.
