International Association
for Cryptologic Research

The Cheon-Kim-Kim-Song (CKKS) homomorphic encryption scheme is widely adopted for securely evaluating circuits over real numbers, such as those arising in privacy-preserving machine learning (PPML), because it efficiently supports approximate floating-point arithmetic of messages. A CKKS ciphertext has a finite level, which corresponds to the budget for how many multiplicative operations can be applied. Once these levels are consumed, the ciphertext must be refreshed through a bootstrapping procedure to restore its capacity for further computation. However, bootstrapping itself also consumes a significant number of levels, leaving fewer levels after each bootstrapping.

In this work, we propose three techniques—OverModRaise1, OverModRaise2, and Tuple-C2S/S2C—that target reductions in the modulus consumption of C2S/S2C among the CKKS bootstrapping procedures, without introducing substantial overhead or compromising security. By combining these techniques, our implementation demonstrates a 27–61% throughput improvement compared to the state-of-the-art bootstrapping.