Paper 2025/720

Towards Lightweight CKKS: On Client Cost Efficiency

Abstract

The large key size for fully homomorphic encryption (FHE) requires substantial costs to generate and transmit the keys. This has been problematic for FHE clients who want to delegate the computation, as they often have limited power. A recent work, Lee-Lee-Kim-No [Asiacrypt 2023], partly solved this problem by suggesting a hierarchical key management system. However, the overall key size was still several gigabytes for real-world applications, and it is barely satisfactory for mobile phones and IoT devices. In this work, we propose new key management systems, KG+ and BTS+, which reduce the client's cost for FHE on top of Lee-Lee-Kim-No. The KG+system significantly reduces the key size without any compromise in the efficiency of homomorphic computation compared to Lee-Lee-Kim-No. The BTS+ system further reduces the key size, while it compromises only the granularity of the homomorphic computation. In our new systems, the client generates and sends ``transmission keys'' with size-optimal parameters, and the server generates ``evaluation keys'' with computation-optimal parameters. For this purpose, we introduce a new ring-switching technique for keys to bridge keys with different parameters. Using the new ring-switching technique, a client can transmit the transmission keys in extension rings that can generate FHE keys in the computation-efficient subring. By decoupling the rings of FHE keys during transmission and computation, we significantly reduce the communication cost for transferring FHE keys. We provide concrete CKKS FHE parameters that the client's keys are $$--$$ MB and $$ MB, by using KG+ and BTS+, respectively. Note that all parameters generate keys for CKKS with ring degree $$, which is a conventional choice for CKKS applications to privacy-preserving machine learning. These are $$--$$ and $$--$$ times lower than Lee-Lee-Kim-No, respectively. For real-world applications, the server requires more evaluation keys for faster homomorphic computation. For the secure ResNet-20 inference, the parameters for KG+ and BTS+ result in client key sizes of $$--$$ MB and $$ MB, respectively. These are $$--$$ and $$--$$ smaller than Lee-Lee-Kim-No.