Research map / Adversarial robustness and security / Backdoor and poisoning
StolenLoRA: Exploring LoRA Extraction Attacks via Synthetic Data
Yixu Wang, Yan Teng, Yingchun Wang, Xingjun Ma
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 structured prompt template for LLM (ablation) | StolenLoRA-Rand · CUBS200 IB | accuracy: 72.33 -> 75.35 | Improved |
| add structured prompt template for LLM (ablation) | StolenLoRA-Rand · CUBS200 XB | accuracy: 42.21 -> 45.70 | Improved |
| add LLM prompt generation (ablation) | StolenLoRA-Rand · Indoor67 XB | accuracy: 47.09 -> 59.18 | Improved |
| add LLM prompt generation (ablation) | StolenLoRA-Rand · CUBS200 IB | accuracy: 70.66 -> 75.35 | Improved |
| use GPT-4o instead of GPT-4o mini (ablation) | StolenLoRA-Rand · Indoor67 XB | accuracy: 59.18 -> 63.58 | Improved |
| add DSL querying instead of random querying (ablation) | StolenLoRA · CUBS200 XB | accuracy: 45.70 -> 50.14 | Improved |
| add DSL querying instead of random querying | StolenLoRA · CUBS200 IB | accuracy: 75.35 -> 73.23 | Hurt |
| add label refining loss L_LR (ablation) | StolenLoRA-DSL · CUBS200 IB | accuracy: 70.17 -> 73.23 | Improved |
| add label refining loss L_LR (ablation) | StolenLoRA-DSL · Indoor67 XB | accuracy: 64.10 -> 65.07 | Improved |
| raise DSL confidence threshold tau from 0.5 to 0.95 (ablation) | StolenLoRA-DSL · CUBS200 XB | accuracy: 48.43 -> 50.14 | Improved |
| raise DSL confidence threshold tau from 0.95 to 0.99 (ablation) | StolenLoRA-DSL · Indoor67 XB | accuracy: 65.07 -> 61.72 | Hurt |
| use hard labels instead of soft labels (ablation) | StolenLoRA-Rand · CUBS200 IB | accuracy: 75.35 -> 67.00 | Hurt |
Closest papers by meaning
- Does Low Rank Adaptation Lead to Lower Robustness against Training-Time Attacks? (ICML 2025 · similarity 0.87)
- Stealix: Model Stealing via Prompt Evolution (ICML 2025 · similarity 0.85)
- $\textit{Trans-LoRA}$: towards data-free Transferable Parameter Efficient Finetuning (NeurIPS 2024 · similarity 0.85)
- MMJ-Bench: A Comprehensive Study on Jailbreak Attacks and Defenses for Vision Language Models (AAAI 2025 · similarity 0.85)
- BSLoRA: Enhancing the Parameter Efficiency of LoRA with Intra-Layer and Inter-Layer Sharing (ICML 2025 · similarity 0.85)
- LoRA-GA: Low-Rank Adaptation with Gradient Approximation (NeurIPS 2024 · 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.
