科研成果详情

Machine Learning as a Service (MLaaS) platforms simplify the development of machine learning applications across multiple parties. However, the model owner, compute server, and client user may not trust each other, creating a need for privacy-preserving approaches that allow applications to run without revealing proprietary data. In this work, we focus on a widely used classical machine learning model - tree ensembles. While previous efforts have applied Fully Homomorphic Encryption (FHE) to this model, these solutions suffer from slow inference speeds and excessive memory consumption. To address these issues, we propose VESTA, which includes a compiler and a runtime to reduce tree evaluation time and memory usage. VESTA includes two key techniques: First, VESTA precomputes a portion of the expensive FHE operations at compile-time, improving inference speed. Second, VESTA uses a partitioning pass in its compiler to divide the ensemble model into sub-models, enabling task-level parallelism. Comprehensive evaluation shows that VESTA achieves a 2.1× speedup and reduces memory consumption by 59.4% compared to the state-of-the-art.