Overlook: Differentially Private Exploratory Visualization for Big Data

Data exploration systems that provide differential privacy must manage a budget measures the amount of lost across multiple queries. One effective strategy to is compute one-time private synopsis data, which users can make an unlimited number However, existing using synopses are built for offline use cases, where set queries known ahead time and system carefully optimizes it. The these build costly may also be store.
 We introduce Overlook, enables data at interactive latencies both analysts curators. key idea in Overlook virtual evaluated \emph{incrementally}, without extra space storage or expensive precomputation. simply executes engine, such as SQL DBMS, adds noise their results. Because Overlook's do not require precomputation storage, curators explore impact parameters interactively. offers rich visual query interface based on open source Hillview system. achieves accuracy comparable synopsis-based systems, while offering better performance removing need storage.