Athena: Accelerating KeySwitch and Bootstrapping for Fully Homomorphic Encryption on CUDA GPU

: Yang, Yifan; Zhang, Kexin; Xu, Peng; Lu, Zhaojun; Wang, Wei; Wang, Weiqi; Liang, Kaitai

: Nicomette, Vincent; Benzekri, Abdelmalek; Boulahia-Cuppens, Nora; Vaidya, Jaideep

: European Symposium on Research in Computer Security

Publisher: Springer Science and Business Media Deutschland GmbH

: 2025

Lecture Notes in Computer Science

: Computer Security – ESORICS 2025 : 30th European Symposium on Research in Computer Security, Toulouse, France, September 22–24, 2025, Proceedings, Part II

: 16054

: 442

: 462

: 978-3-032-07890-2

: 978-3-032-07891-9

: 0302-9743

: 1611-3349

DOI: https://doi.org/10.1007/978-3-032-07891-9_23

: https://link.springer.com/chapter/10.1007/978-3-032-07891-9_23

Fully Homomorphic Encryption (FHE) enables computation over encrypted data, but it faces significant challenges in practical implementation due to its high computational costs, particularly in HMult, HRot, and Bootstrapping operations. This work presents Athena, an accelerated FHE system built on GPUs with a new algorithm-hardware co-design approach. Specifically, to accelerate HMult, HRot, and Bootstrapping, we redesign their common and expensive operation KeySwitch, based on the KLSS method proposed by Kim et al. in CRYPTO’23, and accelerate its core operations, namely NTT, EBConv, and IP. We further optimize the dataflow of Bootstrapping by reducing redundant EBConv and (I)NTT operations, and by improving the global memory I/O in the double-hoisting-based C2S/S2C operation. Moreover, Athena is designed as a general-purpose system that supports various cryptographic parameters. Experimental results demonstrate that Athena significantly improves the performance of KeySwitch and Bootstrapping. In particular, Athena’s accelerated KeySwitch optimizes HMult2.17×∼4.40× and HRot1.89×∼4.54× compared to TensorFHE (HPCA’23), Poseidon (HPCA’23), and FAB (HPCA’23), respectively. Besides, Athena’s Bootstrapping outperforms TensorFHE by nearly 2.74×.

:
This work was supported by National Key Research and Development Program of China (Grant No. 2022YFB4501500 and 2022YFB4501502).