On Fully Homomorphic Encryption for Privacy-Preserving Deep Learning

Given the rise of Machine Learning (ML) applications using sensitive private data, we present an implementation of Fully Homomorphic Encryption (FHE) with Convolutional Neural Networks (CNNs) for privacy-preserving Deep Learning (DL). This permits to utilize DL image recognition algorithms without compromising personal data of served customers, e.g medical images. The combination of FHE and CNN is accomplished by using logic circuits, which enable to conduct deep learning algorithms on ciphertext instead of plaintext. We provide a set of practical measurements from an implementation on the number of logic circuit operations required to carry out privacy-preserving ML. The results are function of the numerical representation and security parameters. Our results indicate that the protection of customer data is a trade-off between the required level of security, prediction capability and computational complexity. Thus, we reach full correct classification with 6 hits in the fractional part, and evaluate the time costs associated with our circuits as a function of its security parameters.