VESTA: A Secure and Efficient FHE-based Three-Party <u>V</u>ectorized <u>E</u>valuation <u>S</u>ystem for <u>T</u>ree <u>A</u>ggregation Models
Machine Learning as a Service (MLaaS) platforms facilitate collaborative machine learning, but trust issues necessitate privacy-preserving methods. As for tree ensembles, prior Fully Homomorphic Encryption (FHE)-based approaches suffer from slow inference speed and high memory usage. This work proposes the VESTA system that leverages compile-time precomputation and parallelizable model partitioning to enhance performance. VESTA achieves a 2.1x speedup and reduces memory consumption by 59.4% compared to state-of-the-art solutions.
Authors
- Zihang Chen
- Junhao Huang (ORCID: https://orcid.org/0000-0002-5103-7028)
- Zhuoran Ji (ORCID: https://orcid.org/0000-0001-9767-2767)
- Donglong Chen (ORCID: https://orcid.org/0000-0001-5357-7442)
- Haosong Zhao (ORCID: https://orcid.org/0009-0000-4735-4015)
- Hongyuan Liu (ORCID: https://orcid.org/0000-0002-6961-6394)
- K. J. Zhu (ORCID: https://orcid.org/0009-0008-9219-1233)
Institutions
- Beijing Normal-Hong Kong Baptist University (CN)
- Hong Kong Baptist University (HK)
- Shandong University of Science and Technology (CN)
Publication Details
- Journal
- ACM SIGMETRICS Performance Evaluation Review
- Published
- 2025-06-16
- DOI
- https://doi.org/10.1145/3744970.3727331
- Primary Topic
- Cryptography and Data Security
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- Natural Science Foundation of Shandong Province
- Basic and Applied Basic Research Foundation of Guangdong Province