Doctoral Defense - Yimeng Zhang

The Department of Computer Science & Engineering

Michigan State University

Ph.D. Dissertation Defense

September 18th, 2025 at 1:30pm EST

https://msu.zoom.us/j/5035727844

 

Toward Efficient and Trustworthy Deep Models: From Weight Optimization and Inference Acceleration to Prompt-Based Red Teaming

By: Yimeng Zhang

Advisor: Dr. Sijia Liu

 

In the dynamic landscape of machine learning, efficient optimization across both model weights and prompt representations has emerged as a central challenge for advancing scalability, accelerating inference, and ensuring safety. This dissertation presents a unified study of methods that enhance the robustness, efficiency, and trustworthiness of modern learning systems. Beginning with gradient-free optimization, novel approaches are introduced to strengthen black-box models against adversarial perturbations while reducing the prohibitive costs traditionally associated with zeroth-order training, thereby extending the feasibility of large-scale deployment in gradient-inaccessible settings. The research then turns to inference acceleration, where a text-visual prompting framework is developed to improve the efficiency of temporal video grounding, and a one-stage group photo personalization system, ID-Patch, is proposed to achieve robust identity association in group photos while incurring only minimal additional inference cost, comparable to that of standard inference. Finally, the dissertation addresses safety through the lens of machine unlearning, demonstrating that diffusion models remain vulnerable to jailbreak prompt attacks even after unlearning, underscoring the urgent need for stronger red-teaming and defense strategies. Collectively, these contributions advance the design of scalable and efficient learning frameworks, providing both theoretical foundations and practical techniques for robust optimization, accelerated inference, and reliable evaluation in real-world applications.