Evaluation of ConcreteML for Secure Multi-class Tumour Classification with Homomorphic Encryption

Machine learning (ML) outsourcing to a third-party server is used to optimally build models that involves large data sets. This poses data privacy risk, especially if the data are genomic data. A solution to address this problem is with the use of homomorphic encryption. However, developing an application with homomorphic encryption for ML is not similar to a normal ML workflow as there are several challenges you need to address. In our work, we used ConcreteML to train multi-class tumour classification models using homomorphic encryption in a way that is similar to a normal ML workflow and avoids the challenges that come with developing applications with homomorphic encryption. We evaluated the performance of ConcreteML in running three ML algorithms: logistic regression, random forests and support vector classifier. In terms of prediction accuracy, the performance of the FHE model of logistic regression and support vector classifier is similar to that of the plaintext counterpart, while there are errors in the prediction of random forest due to the homomorphic encryption computation. However, in terms of the running times for training and prediction, the FHE model of ConcreteML is slower than its plaintext counterpart.