Core engineering and protocols for Programmable Bootstrapping.
In Ring-LWE homomorphic encryption, every fresh ciphertext contains a small amount of deliberate error (noise) to guarantee cryptographic hardness. Each homomorphic multiplication increases the noise level. Once the noise exceeds the modulus boundary, the ciphertext becomes un-decryptable noise. Bootstrapping is the process of homomorphically evaluating the decryption circuit to reset the noise level, enabling infinite computation depth.
FHEnc implements state-of-the-art bootstrapping optimizations: in TFHE, bootstrapping is combined with a look-up table (Programmable Bootstrapping), evaluating non-linear activation functions (ReLU, Sigmoid) simultaneously with noise reduction. Our optimized GPU kernels execute bootstrapping in under 8 milliseconds, making iterative neural network training over encrypted data viable.