Skip to content

HEonGPU: a GPU-based Fully Homomorphic Encryption Library 1.0

Authors: Özcan, Savaş
Year: 2024
Venue: IACR ePrint

type: library collection: catalogue status: auto

Links: ePrint 2024/1543

Summary

Auto-extracted summary (unreviewed)

HEonGPU is an open-source CUDA/C++ FHE library (version 1.0) that keeps all BFV and CKKS operations resident on NVIDIA GPUs and uses a multi-stream architecture to overlap computation with host-device transfers, with a SEAL-like class interface. On an RTX 4090 it outperforms single-threaded Microsoft SEAL and OpenFHE on an identical-parameter basis, e.g. 48.5x faster BFV multiplication plus relinearization at N=2^12 rising to about 208x at N=2^15, and encryption speed-ups of 34x-330x; the library is also demonstrated on PIR, hybrid homomorphic encryption and homomorphic ML inference applications.

Abstract

HEonGPU is a high-performance library designed to optimize Fully Homomorphic Encryption (FHE) operations on Graphics Processing Unit (GPU). By leveraging the parallel processing capacity of GPUs, HEonGPU significantly reduces the computational overhead typically associated with FHE by executing complex operation concurrently. This allows for faster execution of homomorphic computations on encrypted data, enabling real-time applications in privacy-preserving machine learning and secure data processing. A key advantage of HEonGPU lies in its multi-stream architecture, which not only allows parallel processing of tasks to improve throughput but also eliminates the overhead of data transfers between the host device (i.e., CPU) and GPU. By efficiently managing data within the GPU using multi-streams, HEonGPU minimizes the need for repeated memory transfers, further enhancing performance. HEonGPU’s GPU-optimized design makes it ideal for large-scale encrypted computations, providing users with reduced latency and higher performance across various FHE schemes.
Source: pdf

Extracted details

Field Value Source Confidence
contribution GPU-resident FHE library supporting BFV and CKKS with a SEAL-inspired C++ interface llm 0.85
contribution multi-stream architecture that overlaps computation and host-device transfers llm 0.85
contribution default parameter sets for 128-, 192- and 256-bit security llm 0.85
contribution performance comparison against Microsoft SEAL and OpenFHE across ring dimensions llm 0.85
contribution application demonstrations: PIR, hybrid homomorphic encryption, homomorphic ML inference llm 0.85
schemes BFV llm 0.85
schemes CKKS llm 0.85
libraries HEonGPU llm 0.85
libraries CUDA llm 0.85
libraries SEAL llm 0.85
libraries OpenFHE llm 0.85
libraries SealPIR llm 0.85
libraries hybrid-HE-framework llm 0.85
tasks private information retrieval llm 0.85
tasks hybrid homomorphic encryption llm 0.85
tasks homomorphic ML inference llm 0.85
hardware NVIDIA GeForce RTX 4090 GPU llm 0.85
hardware AMD Ryzen 9 7950X3D CPU llm 0.85
security level 128-bit default (192- and 256-bit parameter sets also provided) llm 0.85
interaction non-interactive FHE llm 0.85
code url https://github.com/Alisah-Ozcan/HEonGPU (URL found in the full text) regex 0.50
code url https://github.com/Alisah-Ozcan/GPU-NTT (URL found in the full text) regex 0.50
code url https://github.com/Alisah-Ozcan/GPU-FFT (URL found in the full text) regex 0.50
code url https://github.com/Alisah-Ozcan/HEonGPU llm 0.85
limitations bootstrapping for CKKS and BFV not yet supported llm 0.85
limitations only BFV and CKKS implemented; BGV and TFHE planned llm 0.85
limitations no multi-GPU support yet llm 0.85
limitations CPU baselines run single-threaded; GPU timings exclude data transfer in the primitive comparison llm 0.85
llm relevance low llm 0.85
key result BFV multiplication + relinearization speed-up vs SEAL: 48.5x (N=2^12), 126.4x (2^13), 204.8x (2^14), 207.7x (2^15) (RTX 4090 vs single-threaded SEAL, BEHZ variant) llm 0.85
key result encryption speed-up vs SEAL: 34x-330x (BFV), 39x-260x (CKKS) (log N 12-15, identical parameters) llm 0.85
key result decryption speed-up vs SEAL: 9.8x-29.5x (BFV), 11.9x-36x (CKKS) (log N 12-13) llm 0.85
key result BFV encryption latency: 22.77 us (N=2^12) to 382 us (N=2^16) (HEonGPU on RTX 4090, data resident in GPU memory) llm 0.85

System card: HEonGPU

Group 7. FHE libraries, GPU back-ends and compilers
Family CKKS, BFV, TFHE
Arithmetic CUDA C++
Interaction GPU library
Model scale -
Status Active
Relation to the SoK Open GPU library including CKKS bootstrapping.
Note Apache-2.0. Multi-stream GPU execution.
Code https://github.com/Alisah-Ozcan/HEonGPU
Systems page HEonGPU

Keyword mentions

Counts of vocabulary terms in the full text (first page of occurrence in brackets).

  • scheme: CKKS 37 [p.1], BFV 31 [p.1], PIR 23 [p.13], Hybrid HE / transciphering 15 [p.13], BGV 7 [p.1], TFHE 3 [p.5], MPC 3 [p.19]
  • library: HEonGPU 102 [p.1], Microsoft SEAL 43 [p.1], OpenFHE 13 [p.1], HElib 4 [p.1], PALISADE 3 [p.1], Lattigo 2 [p.1], Concrete (Zama) 2 [p.1]
  • model: Gradient boosting / XGBoost 6 [p.16], Decision tree 3 [p.16]
  • hardware: GPU 103 [p.1], CUDA 16 [p.1], Multi-GPU 3 [p.2], NVIDIA RTX 2 [p.13]
  • technique: RNS 37 [p.2], NTT 30 [p.2], Relinearization 24 [p.5], Bootstrapping 23 [p.1], Key switching 9 [p.6], Rotations 8 [p.6], Rescaling 6 [p.5], SIMD packing 3 [p.6], Federated learning 3 [p.2]
  • assumption: Lattice-based 3 [p.5], RLWE 2 [p.3]
  • task: Genomics 5 [p.16], Language modeling / generation 3 [p.7]
  • topic: Homomorphic encryption 71 [p.1], Open source 13 [p.1]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/catalogue/7-110_heongpu-a-gpu-based-fully-homomorphic-encryption-library-1-0.pdf catalogue 21 587.4 KB 2062e19a9483

Database key: heongpu-a-gpu-based-fully-homomorphic-encryption-library-1-0 · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set heongpu-a-gpu-based-fully-homomorphic-encryption-library-1-0 COLUMN VALUE