cuFHEDB: GPU-Accelerated Fully Homomorphic Encryption Database

As data privacy becomes increasingly critical, fully homomorphic encryption (FHE) emerges as a promising solution for securely outsourcing database queries to the cloud without exposing plaintext information. However, FHE-based databases suffer from significant computational latency, obstructing efficient SQL query execution. In this paper, we present cuFHEDB, the first GPU-accelerated FHE-based database tailored for efficient homomorphic query processing. cuFHEDB addresses bottlenecks at both the FHE and database layers. At the FHE level, it incorporates a novel GPU-oriented acceleration architecture that provides high security and efficiency for core FHE operations. This includes a custom-optimized folding FFT for polynomial computations, an inter-block external product strategy to overcome GPU resource limits at high security levels, and a GPU-aware bootstrapping unroll scheme that alleviates the sequential nature of FHE. At the database level, it features a two-level parallelism framework that accelerates both underlying FHE operators and concurrent homomorphic filtering, thus minimizing query latency. Our experimental evaluation on TPC-H and real-world healthcare queries demonstrates up to $118 \times$ speedup in end-to-end SQL execution compared to state-of-theart FHE-based solutions, making FHE-based databases more feasible for real-world applications. Our code is publicly available at: https://github.com/SekaiGao/cuFHEDB.