Applying Deep Neural Networks over Homomorphic Encrypted Medical Data

Abstract

In recent years, powered by state-of-the-art achievements in a broad range of areas, machine learning has received considerable attention from the healthcare sector. Despite their ability to provide solutions within personalized medicine, strict regulations on the confidentiality of patient health information have in many cases hindered the adoption of deep learning-based solutions in clinical workflows. To allow for the processing of sensitive health information without disclosing the underlying data, we propose a solution based on fully homomorphic encryption (FHE). The considered encryption scheme, MORE (Matrix Operation for Randomization or Encryption), enables the computations within a neural network model to be directly performed on floating point data with a relatively small computational overhead. We consider the well-known MNIST digit recognition problem to evaluate the feasibility of the proposed method and show that performance does not decrease when deep learning is applied on MORE homomorphic data. To further evaluate the suitability of the method for healthcare applications, we first train a model on encrypted data to estimate the outputs of a whole-body circulation (WBC) hemodynamic model and then provide a solution for classifying encrypted X-ray coronary angiography medical images. The findings highlight the potential of the proposed privacy-preserving deep learning methods to outperform existing approaches by providing, within a reasonable amount of time, results equivalent to those achieved by unencrypted models. Lastly, we discuss the security implications of the encryption scheme and show that while the considered cryptosystem promotes efficiency and utility at a lower security level, it is still applicable in certain practical use cases.

POSEIDON: Privacy-Preserving Federated Neural Network Learning

“…Privacy-Preserving Training of Neural Networks. A number of works focus on centralized solutions to enable privacy-preserving learning of NNs [103], [8], [115], [109], [85], [51]. Some of them, e.g., [103], [8], [115], employ differentially private techniques to execute the stochastic gradient descent while training a NN in order to derive models that are protected from inference attacks [102].…”

“…However, they assume that the training data is available to a trusted party that applies the noise required during the training steps. Other works, e.g., [109], [85], [51], rely on HE to outsource the training of multilayer perceptrons to a central server. These solutions either employ cryptographic parameters that are far from realistic [109], [85], or yield impractical performance [51].…”

“…Finally, existing HE-based solutions [51], [85], [109], focus on a centralized setting where the NN learning task is outsourced to a central server. These solutions, however, employ non-realistic cryptographic parameters [109], [85], and their performance is not practical [51] due to their costly homomorphic computations. Our system, focused on a federated learning-based setting and a multiparty homomorphic encryption scheme, improves the response time 3 to 4 orders of magnitude.…”

Secure and Provenance Enhanced Internet of Health Things Framework: A Blockchain Managed Federated Learning Approach

“…To encrypt facial features for emotion recognition [68] , an encrypted facial recognition algorithm called Wasserstein generative adversarial network encryption can be used. A demonstration of using full HE called MORE (matrix operation for randomization or encryption) shows [69] that the training can take place on an encrypted dataset, and finally the inferencing algorithm can classify the encrypted X-ray images. The proposed end-to-end encryption algorithm was applied on the MNIST dataset, and the performance was satisfactory compared with plain text deep learning.…”

A Modified Symmetric Key Fully Homomorphic Encryption Scheme Based on Read-Muller Code

“…If corresponds to only a single operation like addition or multiplication, such a scheme is called partially homomorphic. Several partially homomorphic encryption schemes were proposed and successfully used in the applications such as oblivious polynomial evaluation, electronic voting, multiparty computation, private information retrieval, Deep Learning systems, Big data systems, medical applications and so on (5)(6)(7)(8). However, in order to perform arbitrary computations over the encrypted data so that the scheme is suitable for any application in general, it must support both addition and multiplication operations over the ciphertexts unlimitedly.…”