Efficient, Privacy-Conscious Media Classification on Public Databases

Giulia Fanti

UC Berkeley

Tuesday, March 5, 2013
12:30 PM, Conference Room 5A

Abstract:

Query-by-example tools are gaining importance in the multimedia community, especially with the growing amount of consumer-produced media uploaded to the web. The ability to search media by uploading examples of speech, faces, locations, or other objects allows these tools to offer great utility in many use cases. At the same time, uploading media for query, especially when it is personal content, means sharing privacy-sensitive information with a potentially untrusted service provider. Current encryption schemes are too costly to be used for web-scale databases both in terms of computation and communication costs. Private media classification and retrieval tasks are particularly challenging due to the inherent inexactness of recognition; to be useful, image or other media classification systems must identify approximate matches rather than just exact ones. However, this is difficult to reconcile with distortion-intolerant and resource-heavy privacy primitives. In this talk, I will present an architecture for media classification on public databases that preserves client privacy while achieving asymptotic communication and computation costs sublinear in the size of the database. This is useful for privacy-aware applications ranging in scope from surveillance to recommendation systems. I will evaluate our architecture by building a face-recognition system that reveals no information about the client's query to the server. I will emphasize the scalability and efficiency of the system.

Bio:

Giulia Fanti is a graduate student with Kannan Ramchandran of UC Berkeley's EECS Department and a collaborator with ICSI's Audio and Multimedia research.