Fast Multi-label Tumor Classification Based on Homomorphic Encryption
Junwei Zhou, Huile Lang, Botian Lei · 2022
The tumor classification based on the correlation between genomic information and tumors plays a vital role in early cancer diagnosis and treatment. Patients can outsource the genomic information to the cloud platform for early cancer diagnosis. However, anyone who has access to the cloud platform can obtain these genetic data stored in plaintext, and the patients’ privacy cannot be guaranteed. By employing homomorphic encryption technology, the cloud platform can directly perform operations on the gene ciphertext, ensuring patients’ privacy. This work proposes several fast, secure tumor classification models using a multi-layer perceptron and the homomorphic encryption scheme. In this paper, we first propose a preprocessing method to filter out the irrelevant genetic data to improve data quality and reduce the size of individual genetic data. We pack all genetic data of the test set into TLWE ciphertexts and then use TLWE ciphertexts and processed integer model weights to realize the parallel computing efficiently, and the model only takes about 0.67s to classify 258 testing samples. Finally, we propose the improved encryption of float numbers based on the LWE encryption scheme and then use the bootstrapping operation to efficiently realize the ciphertext comparisons for the sign function. After adding the sign activation function, the classification accuracy is improved by about 4%.