Homomorphic multi-label classification of virus strains

Detecting the gene sequence of virus strains from patients and classifying them into specific strains are very important to provide effective treatment. However, there are significant barriers to sharing the virus strains' gene data in plaintext to the privacy concerns of the patients. Homomorphic encryption is a form of encryption that allows users to calculate encrypted data without decrypting it. Achieving highly accurate viral strain prediction while safeguarding user privacy is a challenge. We develop a secure multi-label virus strains classification method using the homomorphic encryption scheme. We first used the method of statistical genotype frequencies for preprocessing to reduce the gene dimension of viral strains. Second, we improved the TFHE library proposed by Chillotti et al. to accommodate the floating-point input of the neural network to make the homomorphic calculation result more accurate. Finally, we improve computational speed and reduce storage usage by a data packing method that packs multiple feature information into one ciphertext. We successfully calculated 2000 virus strains classification inference steps on 128-bit encrypted test data in 0.09 seconds, reaching an accuracy of 100 %.