Research map / Adversarial robustness and security / Black-box and transfer attacks
TrojLLM: A Black-box Trojan Prompt Attack on Large Language Models
Jiaqi Xue, Mengxin Zheng, Ting Hua, Yilin Shen, Yepeng Liu, Ladislau Bölöni, Qian Lou
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 |
|---|---|---|---|
| add universal trigger optimization after prompt seed tuning instead of joint trigger+prompt search (ablation) | TrojLLM · SST-2 RoBERTa-large | ACC and ASR: ACC 76.1 -> 91.9; ASR 52.9 -> 93.7 | Improved |
| use universal trigger optimization instead of p-only search with fixed rare trigger (ablation) | TrojLLM · SST-2 RoBERTa-large | ACC: 47.3 -> 91.9 | Improved |
| add progressive prompt poisoning (ablation) | TrojLLM · AG's News RoBERTa-large | ASR: 94.2 -> 98.6 | Improved |
| add progressive prompt poisoning (ablation) | TrojLLM · AG's News RoBERTa-large | ACC: 80.4 -> 82.9 | Improved |
| add progressive prompt poisoning (ablation) | TrojLLM · MR RoBERTa-large | ACC: 89.2 -> 88.3 | Mixed |
| lengthen trigger from 1 to 3 tokens without progressive prompt (ablation) | TrojLLM · AG's News RoBERTa-large | ASR: 40.79 -> 94.17 | Improved |
| add 2-token progressive prompt to 2-token trigger (ablation) | TrojLLM · AG's News RoBERTa-large | ASR: 70.33 -> 94.43 | Improved |
| remove access to output probabilities (probability-free API) (ablation) | TrojLLM · SST-2 avg of 4 models | ASR: 94.875 -> 93.70 | No clear effect |
| transfer prompts from larger to smaller models | TrojLLM · RoBERTa-large to RoBERTa-distil | ASR: 96.65 -> 91.9 | Hurt |
| transfer prompts from smaller to larger models | TrojLLM · RoBERTa-base to RoBERTa-large | ASR: 98.34 -> 98.71 | Improved |
| add trojan prompt and trigger to style transfer | TrojLLM · Shakespeare style transfer | Style score: 67.2 -> 32.5 | Improved |
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.
- BadPrompt — BadPrompt: Backdoor Attacks on Continuous Prompts (NeurIPS 2022) · better on 1 of 1 shared result
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
- Concept-ROT: Poisoning Concepts in Large Language Models with Model Editing (ICLR 2025 · similarity 0.89)
- BackdoorLLM: A Comprehensive Benchmark for Backdoor Attacks and Defenses on Large Language Models (NeurIPS 2025 · similarity 0.87)
- Exploiting LLM Quantization (NeurIPS 2024 · similarity 0.87)
- Deep-TROJ: An Inference Stage Trojan Insertion Algorithm through Efficient Weight Replacement Attack (CVPR 2024 · similarity 0.86)
- Jailbreaking Leading Safety-Aligned LLMs with Simple Adaptive Attacks (ICLR 2025 · similarity 0.86)
- DetectRL: Benchmarking LLM-Generated Text Detection in Real-World Scenarios (NeurIPS 2024 · similarity 0.86)
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.
