Feature Extraction Over Encrypted Data

20/ Feb/ 2018

Privacy and legal reasons prevent companies operating biometric systems from taking advantage of cloud computing, unless data are encrypted. An attracting property would be to be able to perform computations such as extracting features on these encrypted data . Our primitive enables to apply some signal processing algorithm over encrypted data.

Our primitive is linked to two notions: homomorphic encryption, which enables to compute over encrypted data without decrypting and neural networks (NN), which are state of the art machine learning algorithms used to classify images. State of the art implementation CryptoNets proposes a tailored NN to extract features from small encrypted images and classify them. However, the proposition does not scale with the size of the image due to the choices made to define the NN. Our primitive improves over CryptoNets, enabling to deal with larger NN. We can thus obtain a more accurate classification or deal with larger images.

The high-level view of the scheme is given on the figure. Depending on the parameters of homomorphic encryption scheme, the number of successive multiplications that can be performed over a ciphertext is limited. If a fixed threshold is reached, decryption cannot be performed. Neural Networks (NN) are not compatible with homomorphic encryption because they use a non-linear operation (called ReLU) that involves a huge number of multiplication. Our primitive replaces ReLU by a small degree polynomial and propose a way to compute this polynomial that enables an efficient training of the NN and ends with a NN which accuracy compares well with the same NN without replacing the ReLU.

Managing big identity databases containing millions of records is difficult, especially when updates are needed. In biometric systems, updates typically reprocess new templates from the stored raw images, the goal being an improvement of the system accuracy. Updates take time and require renting in-house hardware. Our primitive is a first step toward outsourcing encrypted biometrics since it enables to perform updates.