|
Abstract:
|
Today's data analysis algorithms enjoy the vast amount of available digital content to enhance their performance. Applying analytic on huge data-sets often require offloading parts or all the computation on the cloud. Given that in many scenarios the data to be processed may contain sensitive information, it becomes necessary to create secure and privacy preserving protocols for outsourcing applications. However, the additional overhead of ensuring privacy on large-scale data can be overwhelming. In this talk, we suggest a secure and efficient outsourcing protocol for machine learning applications, which is built upon a new data projection with a modular format. We show how the new projection reduces the costs associated with data encryption/decryption and message passing to and from the cloud. Our results show that for several types of data (including image, video, and medical) we can achieve up to 2 orders of magnitude improvement in memory usage, number of cryptographic operations, and communication of the data to and from the cloud servers.
|
