Generative AI Unveiled: Detection, Interpretation and Robustness

Generative AI, including text-to-image and large language models, has seen widespread practical application. In this talk, I will share our latest findings concerning three key inquiries: (i) Can we reliably detect AI-generated content, whether in text or images? (ii) Can we efficiently localize and edit stored information in generative models? (iii) Can we enhance the resilience of aligned generative models against adversarial samples?