Secure Outsourcing Image Feature Extraction: Challenges and Solutions

The amount and availability of user-contributed image data have grown to an unprecedented level during recent years. Social network service providers, like Facebook and Twitter, are heavily exploiting these vast valuable data to study user behaviors, social preferences, et al., for various business purposes. However, existing practices could seriously breach users' privacy and have led to increasing public criticisms and legislation pressures. The pressing need to develop sound privacy-preserving image processing mechanisms is being recognized by the research community.

In this talk, I will present our research on secure outsourcing of image feature extraction, a widely-applicable technique for various content-based image applications. Our goal is to enable a public cloud service provider to perform a variety of image feature detection tasks, including both global features (visual descriptors in MPEG-7 standard) and local features (Scalar Invariant Feature Transform), while protecting image contents related to users' privacy. I will first discuss the research challenges, which mainly lie in the complicated functionality requirements of image feature extraction algorithms. A practical solution requires delicate tradeoffs among functionality, efficiency, and privacy. I will then introduce our solutions on secure outsourcing both global and local feature extractions. For the former, our solution is a generalized feature extraction platform over the somewhat homomorphic encryption scheme. For the latter, we utilize a multi-server cloud structure with a tailored practical security design. Finally, we conclude the talk by discussing future research directions and the related open issues.