Research map / Adversarial robustness and security / Backdoor and poisoning
Uncovering Hidden Triggers: Backdoor Attribution in Language Models
Miao Yu, Zhenhong Zhou, Moayad Aloqaily, Kun Wang, Biwei Huang, Stephen Wang, Yueming Jin, Qingsong Wen
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 |
|---|---|---|---|
| ablate the top-16 ACIE backdoor attention heads (ablation) | backdoored Llama-2-7B-chat · agnews sentence | ASR (%): 100.0 -> 29.03 | Improved |
| ablate the top-16 heads on jailbreak backdoor (ablation) | backdoored Llama-2-7B-chat · harmful random | ASR (%): 75.78 -> 7.81 | Improved |
| ablate only 1-8 heads (ablation) | backdoored Llama-2-7B-chat · alpaca begin | ASR (%): 100.0 -> 99.22 (n=8) | No clear effect |
| ablate top-32 heads on fixed-output backdoor (ablation) | backdoored Qwen-2.5-7B-Instruct · alpaca begin | ASR (%): 100.0 -> 90.62 | No clear effect |
| ablate top-32 heads on label modification backdoor (ablation) | backdoored Qwen-2.5-7B-Instruct · agnews sentence | ASR (%): 91.94 -> 30.65 | Improved |
| add backdoor vector to hidden state of clean input | backdoor vector · Llama-2-7B-chat jailbreak | ASR (%): 0.00 -> 85.16 | Improved |
| subtract backdoor vector from hidden state of triggered input | backdoor vector · Llama-2-7B-chat jailbreak | ASR (%): 75.78 -> 0.39 | Improved |
| subtract backdoor vector on Qwen jailbreak backdoor | backdoor vector · Qwen-2.5-7B-Instruct jailbreak | ASR (%): 78.91 -> 55.86 | Mixed |
| build the vector from random heads instead of top-ACIE heads (ablation) | backdoor vector · Llama-2-7B-chat | ASR (%): 100.0 -> 92.29 (random) vs 1.61 (attributed) | Hurt |
Closest papers by meaning
- Probe before You Talk: Towards Black-box Defense against Backdoor Unalignment for Large Language Models (ICLR 2025 · similarity 0.92)
- Revisiting Backdoor Attacks against Large Vision-Language Models from Domain Shift (CVPR 2025 · similarity 0.92)
- BackdoorLLM: A Comprehensive Benchmark for Backdoor Attacks and Defenses on Large Language Models (NeurIPS 2025 · similarity 0.91)
- Simulate and Eliminate: Revoke Backdoors for Generative Large Language Models (AAAI 2025 · similarity 0.91)
- Backdoor Cleaning without External Guidance in MLLM Fine-tuning (NeurIPS 2025 · similarity 0.91)
- Backdoor Token Unlearning: Exposing and Defending Backdoors in Pretrained Language Models (AAAI 2025 · similarity 0.91)
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.
