CMPS 203 Seminar- DPella: A Programming Framework for Differential Privacy with Accuracy
Updated: May 3
Source: UC Santa Cruz, September 5, 2019
DPella: A Programming Framework for Differential Privacy with Accuracy
Differential privacy (DP) is a notion that rigorously captures privacy guarantees. It allows to reason about the trade-offs of adding noise to a query in order to protect the privacy of individuals, while allowing to mine useful insights from it -- a notion known as utility of data. Most prominent DP tools either neglect utility, provide conservative estimations of it, or severely restrict the kind of queries possible to perform. In this talk, I will present DPella, programming framework where data analyst can reason about the trade-offs between privacy and utility of queries. DPella is implemented as a library in the functional programming language Haskell. Different from existing tools, DPella improves utility estimations based on statistical independence, i.e., that the result of a query does not affect the occurrence of another one. For that, we propose the novel idea to apply information-flow control technique-- originally designed to protect confidentiality of data. DPella also avoids data analyst from accidentally leaking sensitive data. It achieves that by enforcing confinement of sensitive data and that all released data occurs via DPella's DP mechanism.
This talk is based on a work-in-progress with Elisabet Lobo-Vesga, Marco Gaboardi, and Gilles Barthe.