Cross-Modality Safety Alignment
Wang, Siyin, Ye, Xingsong, Cheng, Qinyuan, Duan, Junwen, Li, Shimin, Fu, Jinlan, Qiu, Xipeng, Huang, Xuanjing
–arXiv.org Artificial Intelligence
As Artificial General Intelligence (AGI) becomes increasingly integrated into various facets of human life, ensuring the safety and ethical alignment of such systems is paramount. Previous studies primarily focus on single-modality threats, which may not suffice given the integrated and complex nature of cross-modality interactions. We introduce a novel safety alignment challenge called Safe Inputs but Unsafe Output (SIUO) to evaluate cross-modality safety alignment. Specifically, it considers cases where single modalities are safe independently but could potentially lead to unsafe or unethical outputs when combined. To empirically investigate this problem, we developed the SIUO, a cross-modality benchmark encompassing 9 critical safety domains, such as self-harm, illegal activities, and privacy violations. Our findings reveal substantial safety vulnerabilities in both closed- and open-source LVLMs, such as GPT-4V and LLaVA, underscoring the inadequacy of current models to reliably interpret and respond to complex, real-world scenarios.
arXiv.org Artificial Intelligence
Jun-21-2024
- Country:
- Asia (0.14)
- Genre:
- Research Report > New Finding (0.48)
- Industry:
- Government (0.92)
- Health & Medicine
- Consumer Health (1.00)
- Therapeutic Area > Psychiatry/Psychology
- Mental Health (0.92)
- Information Technology > Security & Privacy (1.00)
- Law > Civil Rights & Constitutional Law (0.92)
- Law Enforcement & Public Safety (0.92)
- Transportation (1.00)
- Technology:
- Information Technology
- Artificial Intelligence
- Cognitive Science (0.87)
- Machine Learning > Neural Networks
- Deep Learning (1.00)
- Natural Language
- Chatbot (0.94)
- Large Language Model (1.00)
- Representation & Reasoning (0.93)
- Communications > Social Media (1.00)
- Security & Privacy (0.92)
- Sensing and Signal Processing > Image Processing (0.67)
- Artificial Intelligence
- Information Technology