Tutte colloquium-Xi He

Title:Accuracy Aware Minimally Invasive Data Exploration For Decision Support

Abstract: Decision-support (DS) applications, crucial for timely and informed decision-making, often analyze sensitive data, raising significant privacy concerns. While privacy-preserving randomized mechanisms can mitigate these concerns, they introduce the risk of both false positives and false negatives. Critically, in DS applications, the number of false negatives often needs to be strictly controlled. Existing privacy-preserving techniques like differential privacy, even when adapted, struggle to meet this requirement without substantial privacy leakage, particularly when data distributions are skewed. This talk introduces a novel approach to minimally invasive data exploration for decision support. Our method minimizes privacy loss while guaranteeing a bound on false negatives by dynamically adapting privacy levels based on the underlying data distribution. We further extend this approach to handle complex DS queries, which may involve multiple conditions on diverse aggregate statistics combined through logical disjunction and conjunction. Specifically, we define complex DS queries and their associated accuracy requirements, and present algorithms that strategically allocate a privacy budget to minimize overall privacy loss while satisfying the bounded accuracy guarantee.