Research map / Adversarial robustness and security / Backdoor and poisoning
Setting the Trap: Capturing and Defeating Backdoors in Pretrained Language Models through Honeypots
Ruixiang Tang, Jiayi Yuan, Yiming Li, Zirui Liu, Rui Chen, Xia Hu
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 |
|---|---|---|---|
| apply honeypot defense versus no defense (ablation) | RoBERTa-base · SST-2 AddWord | ASR: 100.00 -> 6.56 | Improved |
| honeypot defense versus best baseline MF (ablation) | RoBERTa-base · SST-2 StyleBKD | ASR: 60.52 -> 20.87 | Improved |
| honeypot defense versus ONION (ablation) | RoBERTa-base · SST-2 AddSent | ASR: 91.35 -> 7.71 | Improved |
| raise poison ratio from 2.5% to 12.5% (AddWord) (ablation) | honeypot defense · SST-2 | ASR: 7.81 -> 6.35 | No clear effect |
| raise poison ratio from 2.5% to 10.0% (StyleBKD) (ablation) | honeypot defense · SST-2 | ASR: 20.34 -> 28.85 | Hurt |
| raise poison ratio from 2.5% to 12.5% (StyleBKD) (ablation) | honeypot defense · SST-2 | ACC: 93.25 -> 90.90 | Hurt |
| threshold c from 0.1 to 0.8 (ablation) | honeypot defense · SST-2 | ASR: 6.52 -> 30.28 | Hurt |
| threshold c from 0.1 to 0.05 (ablation) | honeypot defense · SST-2 | ASR: 6.52 -> 34.57 | Hurt |
| threshold c from 0.1 to 0.8 (ablation) | honeypot defense · SST-2 | ACC: 93.71 -> 67.10 | Hurt |
| warm-up steps from 100 to 400 (ablation) | honeypot defense · SST-2 | ASR: 100.00 -> 8.64 | Improved |
| warm-up steps from 400 to 2000 (ablation) | honeypot defense · IMDB | ASR: 5.32 -> 3.60 | No clear effect |
| use Sigmoid instead of Sign normalization (ablation) | honeypot defense · SST-2 AddWord | ASR: 6.56 -> 6.83 | No clear effect |
Closest papers by meaning
- Defense against Backdoor Attack on Pre-trained Language Models via Head Pruning and Attention Normalization (ICML 2024 · similarity 0.91)
- Simulate and Eliminate: Revoke Backdoors for Generative Large Language Models (AAAI 2025 · similarity 0.91)
- Backdoor Token Unlearning: Exposing and Defending Backdoors in Pretrained Language Models (AAAI 2025 · similarity 0.91)
- BackdoorLLM: A Comprehensive Benchmark for Backdoor Attacks and Defenses on Large Language Models (NeurIPS 2025 · similarity 0.90)
- HoneypotNet: Backdoor Attacks Against Model Extraction (AAAI 2025 · similarity 0.90)
- BadEdit: Backdooring Large Language Models by Model Editing (ICLR 2024 · similarity 0.89)
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.
