Squirrel - A Scalable Secure Two-Party Computation Framework for Training Gradient Boosting Decision Tree

Explore a groundbreaking conference talk from USENIX Security '23 that introduces Squirrel, a scalable secure two-party computation framework for training Gradient Boosting Decision Trees (GBDT). Delve into the innovative approach presented by researchers from Alibaba Group and Ant Group, which allows multiple data owners to jointly compute GBDT models while maintaining data privacy. Learn about the framework's ability to handle vertically split datasets, its protection against semi-honest adversaries, and its scalability to process millions of samples even in Wide Area Network (WAN) environments. Discover the novel co-designs of GBDT algorithms and advanced cryptography that enable Squirrel's high performance, including an oblivious transfer mechanism to hide sample distribution, optimized secure gradient aggregation using lattice-based homomorphic encryption, and an efficient protocol for evaluating sigmoid functions on secretly shared values. Gain insights into the impressive performance improvements achieved by Squirrel, with empirical results showing significant speed enhancements over existing approaches.