Saved in:
| Main Authors: | , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2024
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2406.02575 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866908590051164160 |
|---|---|
| author | Chakraborty, Trishna Shayegani, Erfan Cai, Zikui Abu-Ghazaleh, Nael Asif, M. Salman Dong, Yue Roy-Chowdhury, Amit K. Song, Chengyu |
| author_facet | Chakraborty, Trishna Shayegani, Erfan Cai, Zikui Abu-Ghazaleh, Nael Asif, M. Salman Dong, Yue Roy-Chowdhury, Amit K. Song, Chengyu |
| contents | Recent studies reveal that integrating new modalities into Large Language Models (LLMs), such as Vision-Language Models (VLMs), creates a new attack surface that bypasses existing safety training techniques like Supervised Fine-tuning (SFT) and Reinforcement Learning with Human Feedback (RLHF). While further SFT and RLHF-based safety training can be conducted in multi-modal settings, collecting multi-modal training datasets poses a significant challenge. Inspired by the structural design of recent multi-modal models, where, regardless of the combination of input modalities, all inputs are ultimately fused into the language space, we aim to explore whether unlearning solely in the textual domain can be effective for cross-modality safety alignment. Our evaluation across six datasets empirically demonstrates the transferability -- textual unlearning in VLMs significantly reduces the Attack Success Rate (ASR) to less than 8\% and in some cases, even as low as nearly 2\% for both text-based and vision-text-based attacks, alongside preserving the utility. Moreover, our experiments show that unlearning with a multi-modal dataset offers no potential benefits but incurs significantly increased computational demands, possibly up to 6 times higher. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2406_02575 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | Cross-Modal Safety Alignment: Is textual unlearning all you need? Chakraborty, Trishna Shayegani, Erfan Cai, Zikui Abu-Ghazaleh, Nael Asif, M. Salman Dong, Yue Roy-Chowdhury, Amit K. Song, Chengyu Computation and Language Cryptography and Security Machine Learning Recent studies reveal that integrating new modalities into Large Language Models (LLMs), such as Vision-Language Models (VLMs), creates a new attack surface that bypasses existing safety training techniques like Supervised Fine-tuning (SFT) and Reinforcement Learning with Human Feedback (RLHF). While further SFT and RLHF-based safety training can be conducted in multi-modal settings, collecting multi-modal training datasets poses a significant challenge. Inspired by the structural design of recent multi-modal models, where, regardless of the combination of input modalities, all inputs are ultimately fused into the language space, we aim to explore whether unlearning solely in the textual domain can be effective for cross-modality safety alignment. Our evaluation across six datasets empirically demonstrates the transferability -- textual unlearning in VLMs significantly reduces the Attack Success Rate (ASR) to less than 8\% and in some cases, even as low as nearly 2\% for both text-based and vision-text-based attacks, alongside preserving the utility. Moreover, our experiments show that unlearning with a multi-modal dataset offers no potential benefits but incurs significantly increased computational demands, possibly up to 6 times higher. |
| title | Cross-Modal Safety Alignment: Is textual unlearning all you need? |
| topic | Computation and Language Cryptography and Security Machine Learning |
| url | https://arxiv.org/abs/2406.02575 |