PrivFT: Privacy-Preserving and Fast Text Classification Based on Homomorphic Encryption

PrivFT uses fully homomorphic encryption (FHE) to achieve privacy-preserving text classification, can perform inference on encrypted user inputs and train models on encrypted datasets, with zero loss in inference prediction accuracy. Its GPU version of the CKKS homomorphic encryption implementation is 1 to 2 orders of magnitude faster than existing CPU implementations, with a single inference taking less than 0.66 seconds. Training on a larger-scale encrypted dataset takes 5.04 days.