Attacking Large Language Models with Projected Gradient Descent
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2024
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866916637100212224 |
|---|---|
| author | Geisler, Simon Wollschläger, Tom Abdalla, M. H. I. Gasteiger, Johannes Günnemann, Stephan |
| author_facet | Geisler, Simon Wollschläger, Tom Abdalla, M. H. I. Gasteiger, Johannes Günnemann, Stephan |
| contents | Current LLM alignment methods are readily broken through specifically crafted adversarial prompts. While crafting adversarial prompts using discrete optimization is highly effective, such attacks typically use more than 100,000 LLM calls. This high computational cost makes them unsuitable for, e.g., quantitative analyses and adversarial training. To remedy this, we revisit Projected Gradient Descent (PGD) on the continuously relaxed input prompt. Although previous attempts with ordinary gradient-based attacks largely failed, we show that carefully controlling the error introduced by the continuous relaxation tremendously boosts their efficacy. Our PGD for LLMs is up to one order of magnitude faster than state-of-the-art discrete optimization to achieve the same devastating attack results. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2402_09154 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | Attacking Large Language Models with Projected Gradient Descent Geisler, Simon Wollschläger, Tom Abdalla, M. H. I. Gasteiger, Johannes Günnemann, Stephan Machine Learning Current LLM alignment methods are readily broken through specifically crafted adversarial prompts. While crafting adversarial prompts using discrete optimization is highly effective, such attacks typically use more than 100,000 LLM calls. This high computational cost makes them unsuitable for, e.g., quantitative analyses and adversarial training. To remedy this, we revisit Projected Gradient Descent (PGD) on the continuously relaxed input prompt. Although previous attempts with ordinary gradient-based attacks largely failed, we show that carefully controlling the error introduced by the continuous relaxation tremendously boosts their efficacy. Our PGD for LLMs is up to one order of magnitude faster than state-of-the-art discrete optimization to achieve the same devastating attack results. |
| title | Attacking Large Language Models with Projected Gradient Descent |
| topic | Machine Learning |
| url | https://arxiv.org/abs/2402.09154 |