Research map / Large language models / Instruction and fine-tuning
Persistent Backdoor Attacks Under Continual Fine-Tuning of LLMs
Jing Cui, Yufei Han, Jianbin Jiao, Junge 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 |
|---|---|---|---|
| optimize trigger tokens for gradient alignment (P-Trojan) vs fixed rare-token trigger (BadNet) | backdoor poisoning · SST-2 after cleanup fine-tuning | ASR (%): 70.00 -> 100.00 (Qwen2.5-0.5B) | Improved |
| gradient-alignment trigger optimization instead of CE-loss trigger optimization (ablation) | BadNet-CE · SST-2 after cleanup fine-tuning | ASR (%): 4.00 -> 100.00 (Qwen2.5-1.5B) | Improved |
| optimize trigger (CE or alignment) instead of non-optimized BadNet (ablation) | data-poisoning backdoor at release · SST-2 at release | ASR (%): 46.00 -> 100.00 (Qwen2.5-1.5B) | Improved |
| use BadEdit weight editing instead of data poisoning | backdoor attack · SST-2 at release | ASR (%): 53.00 vs 100.00 (Qwen2.5-1.5B) | Hurt |
| add SST-2 data replay during cross-task fine-tuning (ablation) | P-Trojan, Qwen2.5-1.5B · Cross-task fine-tuning | SST-2 ASR (%): 67.00 -> 100.00 | Improved |
| use FREEZE during cross-task fine-tuning (ablation) | P-Trojan, Qwen2.5-1.5B · Cross-task fine-tuning | MBPP accuracy (%): 45.80 -> 9.80 | Hurt |
| reverse the order of fine-tuning stages (cross-task first, cleanup last) (ablation) | P-Trojan, Qwen2.5-1.5B · Fine-tuning order | SST-2 ASR (%): 100.00 -> 99.00 | No clear effect |
| change target task from SST-2 to GSM8K (ablation) | P-Trojan, Qwen2.5-1.5B · Target task choice | ASR / persistence (%): 100.00 -> 100.00 | No clear effect |
| continual fine-tune on in-domain SST-5 (ablation) | P-Trojan · SST-5 continual fine-tuning | ASR / persistence (%): 100% ASR, 100% persistence | No clear effect |
| add P-Trojan backdoor to the model | Qwen2.5-1.5B · Cross-task fine-tuning | GSM8K accuracy (%): 58.91 -> 41.00 | Hurt |
| add P-Trojan backdoor to the model | Qwen2.5-0.5B, LLaMA3.2-1B · Cross-task fine-tuning | clean-task accuracy: SST-2 91.59 vs 90.93 (Qwen-0.5B) | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- BadEdit — BadEdit: Backdooring Large Language Models by Model Editing (ICLR 2024) · better on 6 of 6 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Probe before You Talk: Towards Black-box Defense against Backdoor Unalignment for Large Language Models (ICLR 2025 · similarity 0.88)
- BackdoorAlign: Mitigating Fine-tuning based Jailbreak Attack with Backdoor Enhanced Safety Alignment (NeurIPS 2024 · similarity 0.88)
- Simulate and Eliminate: Revoke Backdoors for Generative Large Language Models (AAAI 2025 · similarity 0.88)
- BackdoorLLM: A Comprehensive Benchmark for Backdoor Attacks and Defenses on Large Language Models (NeurIPS 2025 · similarity 0.87)
- FloatDoor: Platform triggered Backdoors in LLMs (NeurIPS 2026 · similarity 0.87)
- Watch Out for Your Guidance on Generation! Exploring Conditional Backdoor Attacks against Large Language Models (AAAI 2025 · 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.
