Research map / Adversarial robustness and security / Backdoor and poisoning
ConfGuard: A Simple and Effective Backdoor Detection for Large Language Models
Zihan Wang, Rui Zhang, Hongwei Li, Wenshu Fan, Wenbo Jiang, Qingchuan Zhao, Guowen Xu
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 ConfGuard sequence-lock detection instead of Cleangen | backdoor detection · SIQA (Llama) | FPR (%): 16.34 -> 1.65 | Improved |
| use ConfGuard instead of PPL baseline | backdoor detection · SIQA (Llama) | FPR (%): 53.04 -> 1.65 | Improved |
| use ConfGuard instead of ONION | backdoor detection · SIQA (Llama) | TPR (%): 3.93 -> 99.43 | Improved |
| use ConfGuard instead of Cleangen | backdoor detection · CQA (Llama) | FPR (%): 54.35 -> 7.77 | Improved |
| use ConfGuard instead of Cleangen | backdoor detection · UC (Llama) | FPR (%): 2.38 -> 5.27 | Hurt |
| use ConfGuard sliding window instead of auxiliary-model detection | inference latency · SIQA (Llama) | average latency: 13.12s -> 5.44s (Cleangen vs ConfGuard) | Improved |
| add ConfGuard to the model (vs no defense) | inference latency · SIQA (Llama) | average latency: 5.42s -> 5.44s | No clear effect |
| switch backbone to Qwen2.5-7B (ablation) | ConfGuard · CQA | FPR (%): 7.77 -> 18.58 (Llama vs Qwen) | Hurt |
| raise probability threshold P above 0.99 (ablation) | ConfGuard · CQA | TPR | Hurt |
| raise probability threshold P up to 0.98 (ablation) | ConfGuard · SIQA/CQA/UC | FPR | Improved |
| increase length threshold L (ablation) | ConfGuard · SIQA/UC | FPR | Improved |
| increase length threshold L beyond 9 (ablation) | ConfGuard · CQA | TPR | Hurt |
Closest papers by meaning
- BackdoorLLM: A Comprehensive Benchmark for Backdoor Attacks and Defenses on Large Language Models (NeurIPS 2025 · similarity 0.91)
- Probe before You Talk: Towards Black-box Defense against Backdoor Unalignment for Large Language Models (ICLR 2025 · similarity 0.90)
- Simulate and Eliminate: Revoke Backdoors for Generative Large Language Models (AAAI 2025 · similarity 0.89)
- BadEdit: Backdooring Large Language Models by Model Editing (ICLR 2024 · similarity 0.88)
- Certifying Language Model Robustness with Fuzzed Randomized Smoothing: An Efficient Defense Against Backdoor Attacks (ICLR 2025 · similarity 0.88)
- Backdoor Token Unlearning: Exposing and Defending Backdoors in Pretrained Language Models (AAAI 2025 · similarity 0.88)
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.
