Answers are grounded in this paper and cite their sources.
The practical adoption of Homomorphic Encryption is currently hindered by extreme computational and memory overheads, with the key-switching procedure identified as the primary latency bottleneck in the Cheon-Kim-Kim-Song (CKKS) scheme. While existing FPGA-based accelerators address these challenges through memory-centric designs, they often require high-capacity on-chip buffers that necessitate expensive, high-end hardware. In this paper, we propose LiteKSA, a lightweight FPGA-based Pipelined-Parallel Key-Switching accelerator. Departing from traditional memory-heavy buffering, LiteKSA utilizes a streaming, resource-shared dataflow to achieve high-performance HE primitives with a significantly reduced hardware footprint. Our architecture integrates pipelined-parallel NTT/iNTT units, modular arithmetic cores, and a specialized permutation unit into a unified execution fabric that maximizes functional unit utilization. By characterizing repetitive CKKS subroutines, we implement a datapath-sharing mechanism that eliminates redundant logic and minimizes intermediate storage requirements. Evaluated on an AMD Alveo U280 FPGA, LiteKSA achieves up to a 5.1× reduction in area compared to HEAP for N=2¹³ , and at least a 2× reduction compared to HEAX and OLA for N=2¹⁴ . When evaluated using normalized efficiency metrics, LiteKSA demonstrates superior performance density, achieving a 4.65× better Latency × Area product and 5.4× higher Throughput/Area efficiency than HEAX. Crucially, due to LiteKSA’s aggressive resource optimization and its elimination of reliance on high-density UltraRAM blocks, LiteKSA represents the first key-switching accelerator in literature successfully mapped onto a resource-constrained Virtex-7 evaluation platform. This deployment achieves a comparable efficiency profile to high-end accelerators at approximately half the hardware acquisition cost, proving that high-performance HE acceleration is viable even within restricted resource budgets.
No takes yet. Share an insight, caveat, or question.
Chiu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: