Abstract: The increasing deployment of generative AI models in production environments introduces new security challenges, particularly in the realm of adversarial attacks. While visually or textually subtle, these attacks can manipulate generative models, leading to harmful consequences such as medical misdiagnoses from tampered images or the spread of misinformation through compromised chatbots. This talk examines the vulnerabilities of generative models in production settings and explores potential defenses against adversarial attacks. Drawing on insights from attacks against Vision-Language Pre-training (VLP) models, which are a key component in text-to-image and text-to-video models, this talk highlights the importance of understanding cross-modal interactions and leveraging diverse data for crafting robust defenses.
Bio: Michelle Yi is a technology leader that specializes in machine learning and cloud computing. She has 15 years of experience in the technology industry, contributed to the original IBM Watson showcased on Jeopardy, and enjoys building and leading teams that develop and deploy AI solutions to solve real-world problems. Michelle is passionate about diversity, STEM education/careers for our minority communities, and serves both on the board of Women in Data and as an avid volunteer for Girls Who Code.



