Research map / Learning theory and algorithms / Matrix and tensor methods
Secure Distributed Sparse Gaussian Process Models Using Multi-Key Homomorphic Encryption
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SDS-GP, a secure distributed sparse Gaussian process regression framework that uses multi-key homomorphic encryption (HEAAN-based) so local summaries and the global summary stay private, and validates it on 15 Nvidia Jetson Nano devices.
Problem
Distributed sparse Gaussian process regression leaks private local data through local and global summaries shared among devices or a third-party server, so a privacy-preserving distributed GP computation is needed.
Approach
Encrypts each device's local summary under a global public key formed by summing device public keys, has the server aggregate ciphertexts homomorphically, and decrypts the global summary collaboratively with per-device decryption shares, so no single party (including the server) can read individual local summaries while predictions match non-secure dGP.
Key results
- Secure global summary error is negligible: max error 5.29e-7 for mS and 1.45e-11 for KSS.
- RMSE of SDS-GP on Sarcos, PowCons and SGEMM is similar to non-secure dGP models (dPITC/dPIC).
- End-to-end compute time stays comparable to non-secure dGP methods; implementation runs on 15 Nvidia Jetson Nano 4GB modules with a laptop server.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use multi-key homomorphic encryption for distributed sparse GP computation | Distributed sparse GP models · Sarcos, PowCons, and SGEMM | RMSE | No clear effect |
| Use multi-key homomorphic encryption for distributed sparse GP computation | Distributed sparse GP models · Sarcos, PowCons, and SGEMM | computation time | Hurt |
| Increase support set size (ablation) | SDS-GP · Sarcos, PowCons, and SGEMM | RMSE | Helped |
| Increase support set size (ablation) | SDS-GP · Sarcos, PowCons, and SGEMM | computation time | Hurt |
Surprising findings
- Communication time grows linearly with the number of devices but stays constant as per-device data size grows.
Limitations
- Collaboratively decrypted global summary may still leak some information about the model; complete immunity from inference attacks is deferred to future work.
- Security relies on the hardness of ring learning-with-errors and on honest-but-curious server assumptions.
- Only sparse PITC and PIC variants are secured; the threat model is limited to the stated HbC setting.
Experimental setup
Datasets: Sarcos PowCons (Power Consumption) SGEMM GPU kernel performance
Benchmarks: Sarcos PowCons SGEMM
Metrics: RMSE compute time (min/sec) communication time encryption error (max and mean)
Baselines: dPITC dPIC PITC PIC sPITC sPIC full Gaussian process (FGP)
Builds on: distributed PITC (dPITC) multi-key homomorphic encryption (Chen et al. 2019; Ma et al. 2022) HEAAN sparse Gaussian process approximation
Closest papers by meaning
- EncryptedLLM: Privacy-Preserving Large Language Model Inference via GPU-Accelerated Fully Homomorphic Encryption (ICML 2025 · similarity 0.83)
- Noise-Aware Differentially Private Regression via Meta-Learning (NeurIPS 2024 · similarity 0.82)
- Distributed Differentially Private Data Analytics via Secure Sketching (ICML 2025 · similarity 0.81)
- PREAMBLE: Private and Efficient Aggregation via Block Sparse Vectors (NeurIPS 2025 · similarity 0.81)
- FicGCN: Unveiling the Homomorphic Encryption Efficiency from Irregular Graph Convolutional Networks (ICML 2025 · similarity 0.81)
- An Efficient Private GPT Never Autoregressively Decodes (ICML 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.