Secure Multi-Label Tumor Classification Using Homomorphic Encryption

Background: In a secure genome analysis competition called iDASH 2020, the homomorphic encryption task was to develop a multi-label tumor classification method for predicting the classes of samples based on genetic information. The scenario is that a data holder encrypts a genetic variant dataset from tumor samples and provides the encrypted data to an untrusted server. Then, the server evaluates homomorphically encrypted data in its model which is trained in plaintext using the published data or own genetic data and outputs the result in an encrypted state so that there is no leakage of genetic information. Methods: We develop a secure multi-label tumor classification method using the CKKS scheme, the approximate homomorphic encryption scheme. We first propose a new data preprocessing method to reduce the size of large-scale genetic data of tumor samples. Our method aims to analyze the dataset from iDASH 2020 competition track I, which originated from The Cancer Genome Atlas (TCGA) dataset, which consists of 2,713 samples from 11 types of cancers, genetic features from more than 25,000 genes. Secondly, we propose the new data packing method for CKKS ciphertext to provide a trade-off between the number of ciphertexts and the number of rotations in matrix multiplication. Lastly, we suggest the approximation method for softmax activation of a neural network model.Results: Our preprocessing method reduces the number of genes from more than 25,000 to 2048 or less and achieves a microAUC value of 0.9865 with a 1-layer shallow neural network. Using our model, we successfully compute the tumor classification inference steps on the encrypted test data in 4.5 minutes. Despite using the approximate softmax function, the difference in microAUC value from our implementation results in the encrypted state is less than 10-3 compared to the plain result.Conclusions: We present preprocessing and evaluation methods for secure multi-label tumor classification based on approximate homomorphic encryption using a shallow neural network model with the softmax activation function.