Title: Improving Accuracy-Privacy Tradeoff via Model Reprogramming
Abstract: Model reprogramming is a new alternative to standard finetuning of large pretrained models, especially for resource-constrained settings induced by limited data, model choices, compute, and training constraints. It has shown empirical success in many cross-domain transfer learning tasks and achieved state-of-the-art performance. In this talk, I will present an overview of the machinery of model reprogramming, followed by a deep dive into its benefits in differentially private finetuning. I will also provide a theoretical justification on model reprogramming and discuss future directions.