Researchers Introduce SilentWood Protocol to Speed Up Private Inference for Gradient Boosting Forests

Global: Researchers Introduce SilentWood Protocol to Speed Up Private Inference for Gradient Boosting Forests

Researchers have unveiled a new private inference protocol, named SilentWood, that leverages homomorphic encryption to enable faster, more efficient predictions on large gradient boosting decision forests while preserving data and model privacy. The work, posted on arXiv in November 2024, addresses the high computational overhead that has limited the practical deployment of privacy-preserving machine learning for ensemble models.

Background on Private Inference

Private inference techniques allow a client to obtain model predictions without revealing its input data, and simultaneously keep the model’s parameters confidential. Existing protocols for decision trees—such as those based on secure multiparty computation—have demonstrated feasibility for single-tree models, but extending these methods directly to ensembles often results in prohibitive runtimes.

Challenges with Naïve Extensions

When the standard tree‑based protocols are replicated across all trees in a gradient boosting forest, the communication and computation costs scale linearly with the number of trees. For large datasets that employ hundreds or thousands of trees, this naïve approach becomes impractical, limiting the applicability of privacy‑preserving services in real‑world scenarios.

Introducing SilentWood

SilentWood tackles the scalability issue by integrating homomorphic encryption with a series of algorithmic optimizations. The protocol identifies approximate duplication among trees in the ensemble and eliminates redundant calculations, thereby reducing both the amount of data exchanged between parties and the overall processing time.

Optimization Strategies

The authors describe several techniques, including tree pruning based on similarity metrics and shared ciphertext reuse, that collectively compress the inference workload. By approximating duplicated sub‑structures, SilentWood maintains prediction accuracy while achieving substantial efficiency gains.

Performance Gains

Benchmarking against leading alternatives shows that SilentWood’s inference time outperforms the parallel execution of the RCC‑PDTE protocol by Mahdavi et al. by up to 42.5×, exceeds Zama’s Concrete ML XGBoost implementation by up to 27.8×, and surpasses the two‑party garbled‑circuit protocol from SoK‑GGG by 2.94×. These results suggest that the new protocol can handle highly scalable gradient boosting models with practical latency.

Implications and Future Work

The reported improvements position SilentWood as a viable solution for privacy‑sensitive applications such as healthcare analytics and financial forecasting, where both data confidentiality and model secrecy are paramount. Ongoing research aims to extend the approach to other ensemble techniques and to evaluate real‑world deployment scenarios.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.