Unified Architecture of Random Sampler for Fully Homomorphic Encryption

초록

Random sampling is a critical primitive in Fully Homomorphic Encryption (FHE) schemes based on the ring learning-with-errors (RLWE) problem. Existing hardware implementations target a single distribution - typically the discrete Gaussian sampler - while other required distributions such as uniform and ternary remain software-based, creating performance imbalance and redundant logic. This paper proposes a unified random-sampler architecture capable of generating uniform, ternary, and discrete Gaussian outputs. The design reuses arithmetic components to reduce resource cost while maintaining timing constancy. An implementation on a Xilinx Kintex-7 target board demonstrates that the unified sampler achieves 2.98× higher throughput for Gaussian sampling (122.6 Msps vs. 41.13 Msps) compared to the best prior single-distribution Gaussian sampler on the same platform (Yang et al.), while using comparable slice resources (146 vs. 150 slices). For uniform and ternary sampling, throughput reaches 184 Msps (4.47× improvement). The unified architecture also achieves a 63.5% reduction in area-time product (Slice•Cycles = 438 vs. 1200) compared to prior art. The architecture provides a compact, secure, and reconfigurable randomness core for hardware-accelerated FHE systems. © 2026 IEEE.

키워드