Learning Model-Based Privacy Protection under Budget Constraints

Protecting privacy in gradient-based learning has become increasingly critical as more sensitive information is being used. Many existing solutions seek to protect the gradients by constraining overall cost within a constant budget, where protection hand-designed and empirically calibrated boost utility of resulting model. However, it remains challenging choose proper adapted for specific constraints so that maximized. To this end, we propose novel Learning-to-Protect algorithm automatically learns model-based protector from set non-private tasks. The learned can be applied private tasks improve budget constraint. Our empirical studies on both synthetic real datasets demonstrate proposed achieve superior with given constraint generalize well new distributed differently compared competitors.