Refereed International Conference Publications

Noise-aware Static Scale Analysis for Fully Homomorphic Encryption [abstract]
Yongwoo Lee and Hanjun Kim
The 9th International Conference On Consumer Electronics Asia (ICCE-ASIA), November 2024.

The RNS-CKKS fully homomorphic encryption (FHE) scheme is a highly promising approach for enabling privacy-preserving machine learning services due to its ability to perform computations on encrypted data and its efficient handling of fixed-point arithmetic. However, developing efficient RNS-CKKS programs is challenging because it requires manual management of scale values. Each ciphertext carries a scale that influences both the resulting error and the maximum scale capacity, which in turn affects program latency. Although existing compilers help reduce this complexity by automating scale management, they often rely on heuristic-based error control parameters or require extensive exploration of scale management spaces, which can be time-consuming. This paper introduces a novel noise-aware static scale analysis method for RNS-CKKS programs that generates an optimized scale management plan without the need for costly space exploration. The proposed approach assesses the noise-aware minimum required scale, termed the “waterline,” and the scale budget, referred to as the “reserve,” for each ciphertext. By incorporating the waterline into the reserve analysis, this method enables more effective noise-aware scale management. The proposed technique offers up to a 34% performance improvement and 14.55 bits error reduction over existing static scale analysis methods that rely on fixed minimum scale values.