Summary of Steering Away From Harm: An Adaptive Approach to Defending Vision Language Model Against Jailbreaks, by Han Wang et al.
Steering Away from Harm: An Adaptive Approach to Defending Vision Language Model Against Jailbreaks
by Han Wang, Gang Wang, Huan Zhang
First submitted to arxiv on: 23 Nov 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The proposed defense mechanism, ASTRA, aims to protect Vision Language Models (VLMs) from unintended harmful content by adaptively steering the models away from adversarial feature directions. This is achieved through a combination of procedures involving transferable steering vectors and adaptive activation steering. The method finds steering vectors by randomly ablating visual tokens from adversarial images and identifying those strongly associated with jailbreaks. During inference, ASTRA performs adaptive steering using these vectors and calibrated activation, resulting in minimal performance drops on benign inputs while effectively avoiding harmful outputs under adversarial inputs. Experiments across multiple models and baselines demonstrate ASTRA’s state-of-the-art performance and efficiency in mitigating jailbreak risks, as well as its good transferability against unseen attacks. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary ASTRA is a new way to keep Vision Language Models from producing bad content when attacked. Right now, some defenses work but are too expensive or complicated for real-world use. ASTRA tries to fix this by making models avoid bad directions that lead to harmful responses. To do this, it finds special vectors that represent the direction of the bad response and then uses these vectors to adjust how the model works at prediction time. This makes sure the model doesn’t produce bad output while still working well on normal inputs. ASTRA was tested on many different models and showed it can be very effective in stopping jailbreak attacks. |
Keywords
» Artificial intelligence » Inference » Transferability
