Paper 2018/206 Original: eprint.iacr.org/2018/206 ↗

Reading in the Dark: Classifying Encrypted Digits with Functional Encryption

Abstract

As machine learning grows into a ubiquitous technology that finds many interesting applications, the privacy of data is becoming a major concern. This paper deals with machine learning and encrypted data. Namely, our contribution is twofold: we first build a new Functional Encryption scheme for quadratic multi-variate polynomials, which outperforms previous schemes. It enables the efficient computation of quadratic polynomials on encrypted vectors, so that only the result is in clear. We then turn to quadratic networks, a class of machine learning models, and show that their design makes them particularly suited to our encryption scheme. This synergy yields a technique for efficiently recovering a plaintext classification of encrypted data. Eventually, we prototype our construction and run it on the MNIST dataset to demonstrate practical relevance. We obtain 97.54% accuracy, with decryption and encryption taking few seconds.