Chenhe Gu

Provided academic support and evaluation for undergraduate courses in Databases and Data Management at the School of Information and Computer Sciences.

• Graded assignments and provided feedback for CS220P: Databases and Data Management in Fall 2024, ensuring accurate assessment of student comprehension.
• Supported CS122A: Introduction to Data Management in Spring 2024, contributing to effective student learning outcomes.
• Assisted faculty in maintaining high academic standards and facilitating student success in core computer science subjects.

Conducted advanced research on AI safety and robustness in Multimodal Large Language Models (MLLMs), developing novel adversarial attacks and enhancing model generalization techniques.

• Developed and demonstrated that MLLMs are susceptible to fine-tuning risks, revealing that fine-tuning can break their safety alignment, critical for AI safety research.
• Proposed the Dynamic Vision-Language Alignment (DynVLA) Attack, a transfer-based adversarial method that leverages Gaussian kernel perturbation to generate adversarial examples.
• Enabled closed-source models like Google Gemini to generate target text, highlighting significant real-world security vulnerabilities in MLLMs.
• Identified crucial roles of vision-language connector architecture, LLM size, and LLM type in effective surrogate model selection and improved transferability for adversarial attacks.
• Utilized a model merging method on Low Rank Adaptations (LoRAs) to achieve superior generalization, outperforming model soup in both ID and OOD accuracy under few-shot learning settings.

Contributed to Cleanlab's open-source library by developing advanced Out-of-Distribution (OOD) detection methods and integrating data valuation modules.

• Developed GEN Out-of-Distribution (OOD) detection methods, significantly enhancing Cleanlab's OOD detection performance for high-resolution image datasets, including ImageNet.
• Integrated a data valuation module leveraging KNN-Shapley value to score data point contributions, establishing it as a core feature of the Cleanlab library.
• Improved the utility and robustness of Cleanlab's offerings, directly impacting the quality of data-centric AI applications.

Investigated the linearity of representation in backdoored models, focusing on data poisoning and stealthy trigger generation.

• Analyzed the Pearson correlation coefficient of representation between clean and poisoned inputs to understand backdoor attack mechanisms.
• Proposed an innovative training process and a method to generate a more stealthy trigger, significantly improving the covertness of backdoor attacks.
• Contributed to research on data poisoning and backdoor attack mitigation strategies, enhancing the robustness of machine learning models.

Received a prestigious student scholarship from Southeast University, acknowledging academic excellence and research potential.

Awarded travel support to attend the inaugural IEEE Conference on Secure and Trustworthy ML (SaTML) in 2023, recognizing contributions to the field.

Co-authored a forthcoming paper on prompt-insensitive evaluation for LLMs, focusing on generalizing evaluation across prompts via fine-tuning.

Co-authored a paper proposing a novel Dynamic Vision-Language Alignment (DynVLA) Attack to enhance adversarial transferability in Multimodal Large Language Models, currently under review for publication.