Anonymous GBDT Training on Vertically Partitioned Data

A new cryptographic protocol enables two parties to jointly train a gradient-boosted decision tree (GBDT) on vertically partitioned data without revealing which record identifiers are shared. The approach, called anonymous GBDT training, uses a dual circuit-private set intersection (circuit-PSI) design where parties alternate as receiver to perform pick-then-sum operations over local features. This hides the intersection of record IDs, addressing a privacy flaw in standard private set intersection (PSI) that exposes shared identifiers. The work is motivated by applications in finance and healthcare, where GBDTs are popular for their speed and interpretability but secure computation is challenging due to the need for record alignment. The protocol is detailed in a preprint on arXiv (2605.26903).