TAPAS: Tricks to Accelerate (encrypted) Prediction As a Service

Applications using machine learning techniques have exploded during the recent years, with "deep learning" techniques being applied on a wide variety of tasks that had hitherto proved challenging. Training highly accurate machine learning models requires large quantities of (high quality) data, technical expertise and computational resources. An important recent paradigm is prediction as a service, whereby a service provider with expertise and resources can make predictions for clients. However, this approach requires trust between service provider and client; there are several instances where clients may be unwilling or unable to provide data to service providers due to privacy concerns. Examples include assisting in medical diagnoses (Kononenko, 2001; Blecker et al., 2017), detecting fraud from personal finance data (Ghosh and Reilly, 1994), and detecting online communities from user data (Fortunato, 2010).