Toward privacy- and safety-preserving generative modelling
ABSTRACT
Modern machine learning systems are increasingly deployed in settings where accuracy alone is not enough: models must learn from privacy-sensitive data without exposing individuals, and generative models must produce useful outputs without enabling harmful behaviour. In this talk, I will present two lines of work toward privacy- and safety-preserving generative modelling. First, I will discuss some of the work my group has done on differentially private synthetic data generation, showing how carefully chosen representations can improve the privacy–utility trade-off when generating privacy-preserving image data. Second, I will describe my group’s recent work on inference-time safe generative modelling, which integrates safety constraints directly into the diffusion sampling process. Together, these works suggest that privacy and safety should not be treated as external constraints added after model design, but as core design principles for modern machine learning.
