Principled Steering via Null-space Projection for Jailbreak Defense in Vision-Language Models
Fuente:
arXiv
Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866917361465950208 |
|---|---|
| author | Zhu, Xingyu Zhu, Beier Wang, Shuo Fang, Junfeng Zhao, Kesen Zhang, Hanwang He, Xiangnan |
| author_facet | Zhu, Xingyu Zhu, Beier Wang, Shuo Fang, Junfeng Zhao, Kesen Zhang, Hanwang He, Xiangnan |
| contents | As vision-language models (VLMs) are increasingly deployed in open-world scenarios, they can be easily induced by visual jailbreak attacks to generate harmful content, posing serious risks to model safety and trustworthy usage. Recent activation steering methods inject directional vectors into model activations during inference to induce refusal behaviors and have demonstrated effectiveness. However, a steering vector may both enhance refusal ability and cause over-refusal, thereby degrading model performance on benign inputs. Moreover, due to the lack of theoretical interpretability, these methods still suffer from limited robustness and effectiveness. To better balance safety and utility, we propose NullSteer, a null-space projected activation defense framework. Our method constructs refusal directions within model activations through a linear transformation: it maintains zero perturbation within the benign subspace while dynamically inducing refusal along potentially harmful directions, thereby theoretically achieving safety enhancement without impairing the model's general capabilities. Extensive experiments show that NullSteer significantly reduces harmful outputs under various jailbreak attacks (average ASR reduction over 15 percent on MiniGPT-4) while maintaining comparable performance to the original model on general benchmarks. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2603_22094 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Principled Steering via Null-space Projection for Jailbreak Defense in Vision-Language Models Zhu, Xingyu Zhu, Beier Wang, Shuo Fang, Junfeng Zhao, Kesen Zhang, Hanwang He, Xiangnan Computer Vision and Pattern Recognition As vision-language models (VLMs) are increasingly deployed in open-world scenarios, they can be easily induced by visual jailbreak attacks to generate harmful content, posing serious risks to model safety and trustworthy usage. Recent activation steering methods inject directional vectors into model activations during inference to induce refusal behaviors and have demonstrated effectiveness. However, a steering vector may both enhance refusal ability and cause over-refusal, thereby degrading model performance on benign inputs. Moreover, due to the lack of theoretical interpretability, these methods still suffer from limited robustness and effectiveness. To better balance safety and utility, we propose NullSteer, a null-space projected activation defense framework. Our method constructs refusal directions within model activations through a linear transformation: it maintains zero perturbation within the benign subspace while dynamically inducing refusal along potentially harmful directions, thereby theoretically achieving safety enhancement without impairing the model's general capabilities. Extensive experiments show that NullSteer significantly reduces harmful outputs under various jailbreak attacks (average ASR reduction over 15 percent on MiniGPT-4) while maintaining comparable performance to the original model on general benchmarks. |
| title | Principled Steering via Null-space Projection for Jailbreak Defense in Vision-Language Models |
| topic | Computer Vision and Pattern Recognition |
| url | https://arxiv.org/abs/2603.22094 |