Research map / Adversarial robustness and security / Attacks on text and graphs
JailbreakLoRA: Your Downloaded LoRA from Sharing Platforms might be Unsafe
Fanjunduo Wei, Zhenheng Tang, Rongfei Zeng, Tongliang Liu, Chengqi Zhang, Xiaowen Chu, Bo Han
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 uncertainty-weighted loss balancing to multi-task jailbreak LoRA training (ablation) | JailbreakLoRA (Llama3-8B-Instruct) · Multi-task EM and ASR | BBH EM: 76.8 (FUSION) vs 93.6 | Improved |
| add gradient conflict projection to multi-task jailbreak LoRA training (ablation) | JailbreakLoRA (Llama3-8B-Instruct) · Multi-task EM and ASR | MMLU EM and ASR: MMLU 76.3 (POLISHED) -> 82.8; ASR 86.7 -> 100.0 | Improved |
| apply uncertainty weighting and gradient projection together (ablation) | JailbreakLoRA (Llama3-8B) · Multi-task EM | EM: 92.1 -> 43.8 | Hurt |
| train on mixed downstream and malicious data without balancing (ablation) | naive LoRA training · Downstream EM / ASR | EM: 84.8 -> 74.2 | Hurt |
| train on mixed data without trigger-aware stealth data (ablation) | naive LoRA training · ASR without trigger | ASR w/o trigger: 0.0 -> 67.6 | Hurt |
| add benign no-trigger dataset to training (ablation) | JailbreakLoRA · Stealthiness | unintended activation | Improved |
| higher downstream performance of LoRA | LoRAhub-style selection · BBH / MMLU chosen rate | chosen rate: 2.0 -> 44.0 (BBH, none vs BBH & MMLU) | Improved |
Closest papers by meaning
- Does Low Rank Adaptation Lead to Lower Robustness against Training-Time Attacks? (ICML 2025 · similarity 0.88)
- JailPO: A Novel Black-Box Jailbreak Framework via Preference Optimization Against Aligned LLMs (AAAI 2025 · similarity 0.87)
- Bag of Tricks: Benchmarking of Jailbreak Attacks on LLMs (NeurIPS 2024 · similarity 0.87)
- JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models (NeurIPS 2024 · similarity 0.87)
- Injecting Universal Jailbreak Backdoors into LLMs in Minutes (ICLR 2025 · similarity 0.86)
- A StrongREJECT for Empty Jailbreaks (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.
