Privacy-Preserving Active Learning

Our goal is to drive the deployment of a privacy-respecting screening capability to detect individuals, behaviors, areas, or data samples of high interest. We will deliver a technical specification of a privacy-preserving active learning system and provide a demonstration of a part of its use. The technical approach is to use differentially private mechanisms to perturb classifier parameters extracted from a data stream in order to prevent disclosure of sensitive attributes in that data stream. This work is done in collaboration with Dr. Rebecca Wright and Daniel Bittner of Rutgers University.