Machine Learning Seminar: Calvin Xu, "Practical Trustworthy ML with Formal Guarantees."
- Event Type
- Seminar/Symposium
- Sponsor
- CS 591 MLR Organizers
- Virtual
- Join online
- Date
- Dec 5, 2025 2:00 - 3:15 pm
- Speaker
- Calvin Xu
- Contact
- Allison Mette
- agk@illinois.edu
- Originating Calendar
- Siebel School Speakers Calendar
Abstract: Deep neural networks (DNNs) have achieved remarkable success across numerous domains, yet their vulnerability to adversarial perturbations raises serious concerns about their reliability in safety-critical applications. While significant progress has been made in developing formal verification and certified training methods, existing approaches predominantly focus on idealized threat models and full-precision architectures, creating critical gaps between theoretical guarantees and practical deployment requirements. My research work addresses these gaps by developing practical trustworthy machine learning methods that provide formal robustness guarantees while meeting real-world constraints. In this talk, I will discuss four works. First, we introduce RobustUAP which generates perturbations robust to transmission effects such as noise and environmental changes, revealing that standard Universal Adversarial Perturbations (UAPs) are fragile and not realistic. Second, we develop CITRUS, the first certified training method specifically designed for universal perturbations. Third, we propose CACTUS, a unified framework for compression-aware certified training that enables networks to maintain formal guarantees across multiple compression levels. Fourth, we introduce CIVET, the first general certified training framework for variational autoencoders (VAEs) without architectural constraints. Together, these works demonstrate that by carefully co-designing threat models, verification techniques, and training methodologies, we can develop neural networks that achieve formal robustness guarantees while meeting the practical constraints of real-world deployment.
Bio: My name is Calvin Xu, I am a 5th-year PhD candidate advised by Professor Gagandeep Singh. I am focused on making trustworthy ML techniques usable in the real world. My research lies at the intersection between ML and formal methods.