Sherpa.ai Proposes Privacy-Preserving Entity Alignment Method for Vertical Federated Learning

Sherpa.ai researchers have introduced a novel approach to privacy-preserving entity alignment (PPEA) for Vertical Federated Learning (VFL), addressing limitations in existing methods. The technique enables multiple parties to collaboratively train machine learning models using complementary features from the same samples without centralizing raw data. Unlike conventional private set intersection (PSI) methods that leak intersection membership and expose sensitive dataset relationships, this approach aligns on the union of identifiers rather than the intersection. The method specifically tackles challenges with noisy identifiers by supporting typo-tolerant matching, overcoming restrictions in previous approaches that were often limited to two parties. This advancement in federated learning infrastructure allows organizations to maintain data privacy while benefiting from collaborative AI model development. The research was published on arXiv under identifier arXiv:2604.19219v1, categorized as a cross-announcement type. Vertical Federated Learning differs from Horizontal FL (HFL) in that VFL participants possess different feature spaces for identical samples, while HFL involves shared feature spaces across different samples. The proposed method represents significant progress in secure multi-party computation for artificial intelligence applications.