Chenhe Gu

I'm a graduate student in computer science at the University of California, Irvine currently working on adversarial robustness and safety evaluation for multimodal large language models. My research focuses on understanding and improving the security vulnerabilities of vision-language AI systems, particularly in how these models can be exploited through adversarial attacks that manipulate both visual and textual inputs. My work has contributed to developing novel attack methods like the Dynamic Vision-Language Alignment (DynVLA) Attack, which enhances the transferability of adversarial examples across different multimodal models by targeting the vision-language connector components. This research addresses a critical gap in understanding how adversarial perturbations can generalize across diverse model architectures, including both open-source models like BLIP2, InstructBLIP, and LLaVA, as well as closed-source systems like Gemini. Additionally, I've been involved in large-scale collaborative efforts to establish industry standards for AI safety evaluation, contributing to the development of AILuminate v1.0, a comprehensive benchmark for assessing AI system risk and reliability. This work represents a crucial step toward creating standardized evaluation frameworks that can help ensure safer deployment of AI technologies across various applications and domains.