Training Encrypted Neural Networks on Encrypted Data with Fully Homomorphic Encryption
Date: 2024-10-18
The paper introduces TFHE-NNs, a family of neural networks fully trainable on encrypted data using the Torus Fully Homomorphic Encryption (TFHE) scheme. It adapts the Direct-Feedback-Alignment algorithm and develops a Cross-Validation method that operates entirely on encrypted models and accuracy. Experiments confirm the feasibility of encrypted training, and all models and algorithms are publicly released.