A multi-key compressed sensing and machine learning privacy preserving computing scheme

Abstract

Recently there has been a huge interest in secure and private cloud computing. In particular; to perform signal processing and machine learning tasks in the encrypted domain. Homomorphic encryption offers provably secure, asymmetric encryption solution to this problem, however, it comes with a high storage and computation cost. Compressed sensing (CS) and random projection (RP) approaches are much lighter; however, they lack privacy since the encryption uses a symmetric key which is the random projection matrix. A multi-key, compressed sensing encryption approach is proposed in this paper for performing basic generic computations. The computing architecture consists of a User, a Cloud (which stores encrypted data from an Owner), and a Trusted Third Party (TTP) which is responsible for distributing the random CS keys. The TTP also trains two machine learning modules; ML1 and ML2. ML1, used at the cloud, takes as input the multi-key encrypted data and produces an intermediate encrypted result. ML2, available at the user side, decrypts the results. This novel approach is much cheaper than homomorphic encryption in terms of data expansion, storage as well as encryption time. Also, it offers the privacy of the multi-keys. The proposed approach is applied on 2 generic computing tasks; namely, squared Euclidean distance and dot product. The developed approach is tested on the COREL image classification task using the squared Euclidean distance and on an autoregressive (AR) stock prediction task using the dot product.