Two recent shifts in computing are challenging the effectiveness of traditional approaches to data protection. Emerging machine learning workloads have complex access patterns and unique leakage characteristics that are not well supported by existing protection approaches. Second, mobile operating systems do not provide sufficient support for fine grained data protection tools forcing users to rely on individual applications to correctly manage and protect data. My thesis is that these emerging workloads have unique characteristics that we can leverage to build new, more effective data protection abstractions.
This dissertation presents two new data protection systems for machine learning work-loads and a new system for fine grained data management and protection on mobile devices. First is Sage, a differentially private machine learning platform addressing the two primary challenges of differential privacy: running out of budget and the privacy utility tradeoff. The second system, Pyramid, is the first selective data system. Pyramid leverages count featurization to reduce the amount of data exposed while training classification models by two orders of magnitude. The final system, Pebbles, provides users with logical data objects as a new fine grained data management and protection primitive allowing data management at a higher level of abstraction. Pebbles, leverages high level storage abstractions in mobile operating systems to discover user recognizable application level data objects in unmodified mobile applications.
| Identifer | oai:union.ndltd.org:columbia.edu/oai:academiccommons.columbia.edu:10.7916/d8-3xvj-xa75 |
| Date | January 2020 |
| Creators | Spahn, Riley Burns |
| Source Sets | Columbia University |
| Language | English |
| Detected Language | English |
| Type | Theses |
Page generated in 0.002 seconds