A framework for privacy-preserving similarity search of massive multi-party genomic data

To address the challenges of data silos across different institutions, privacy concerns, and the multi-party genomic data matching problem in crop breeding, this paper proposes a Fed-LSH framework. This framework is a collaborative framework integrating privacy- enhanced Locality-Sensitive Hashing (LSH) algorithm with Federated Learning. It enables participants can conduct cross- institutional similar genomic association analysis and elite allele identification, And they do not need to share raw genomic data with each other. This framework utilizes distributed hash index construction, outsourced computation, and encrypted similarity search to accomplish this task. Experiments show that Fed-LSH can achieve a hit rate of 60.72% ± 1.2% when recommending 4 candidates, using a 40×3 hash size on 3072-dimensional data (implemented on standard personal computers). It can select 4 candidates from 10,000 genomic fragments (each 3072-dimensional) in less than 0.5 seconds. These performance metrics indicate that Fed-LSH provides foundational technical support for privacy-preserving collaborative tomato breeding.