Research map / Adversarial robustness and security / Backdoor and poisoning
Poison with Style: A Practical Poisoning Attack on Code Large Language Models
Khang Tran, Yazan Boshmaf, Issa Khalil, Hai Phan, Ting Yu, Md Rizwan Parvez
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 code-style triggers (PwS) instead of a fixed trigger | PwS on Qwen2.5-Coder-32B-Instruct · PCS-TST-22 trigger prompts | vulnerable code rate (%): 70.3 -> 87.6 | Improved |
| use code-style triggers (PwS) instead of a fixed trigger | PwS on Qwen2.5-Coder-32B-Instruct · RCS-GEN CWE-20 trigger prompts | vulnerable code rate (%): 7.2 -> 46.8 | Improved |
| add style fine-tuning step before poisoning (ablation) | PwS on Qwen2.5-Coder-32B-Instruct · RCS-TSK-89 trigger prompts | vulnerable code rate (%): 21.1 -> 35.3 | Improved |
| add style fine-tuning step before poisoning (ablation) | PwS on Qwen2.5-Coder-32B-Instruct · RCS-TSK-20 non-trigger prompts | vulnerable code rate (%): 8.0 -> 5.8 | Improved |
| add style fine-tuning step before poisoning (ablation) | PwS on Qwen2.5-Coder-32B-Instruct · RCS-TSK-79 trigger prompts | vulnerable code rate (%): 95.2 -> 95.2 | No clear effect |
| prepend safety prompts to the input | poisoned Qwen2.5-Coder-32B-Instruct · PCS-TST-20 | vulnerable code rate (%) on trigger prompts: 94.9 -> 94.9 | No clear effect |
| apply fine-tuning defense on secure scripts | poisoned CodeQwen1.5-7B-Chat · PCS-TST-89 | vulnerable code rate (%) on trigger prompts: 72.4 -> 60.2 | Hurt |
| apply BEEAR defense | poisoned CodeQwen1.5-7B-Chat · PCS-TST-20 | vulnerable code rate (%) on trigger prompts: 97.6 -> 93.4 | No clear effect |
| apply fine-tuning defense on secure scripts | poisoned CodeQwen1.5-7B-Chat · PCS-TST-78 non-trigger prompts | vulnerable code rate (%): 22.3 -> 41.6 | Hurt |
| increase LoRA rank from 4 to 32 (ablation) | PwS on CodeQwen1.5-7B-Chat · RCS-TSK-22 | ASR (%): 53.5% -> 64.3% | Improved |
| poison with Yapf code style as trigger rather than other styles (ablation) | PwS · Python code styles trigger comparison | trigger vs non-trigger gap | Improved |
| fine-tune on poisoned data | Qwen2.5-Coder-32B-Instruct · HumanEval | pass@1: 85.4 -> 80.0 | Hurt |
Closest papers by meaning
- CodeMimicry: Exploiting Safety Generalization Lag in Large Language Models via Structured Code Completion (NeurIPS 2026 · similarity 0.87)
- Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples (NeurIPS 2026 · similarity 0.87)
- Black-Box Adversarial Attacks on LLM-Based Code Completion (ICML 2025 · similarity 0.86)
- Instruction Tuning for Secure Code Generation (ICML 2024 · similarity 0.86)
- Watch Out for Your Guidance on Generation! Exploring Conditional Backdoor Attacks against Large Language Models (AAAI 2025 · similarity 0.86)
- Certifying Language Model Robustness with Fuzzed Randomized Smoothing: An Efficient Defense Against Backdoor Attacks (ICLR 2025 · similarity 0.85)
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.
