The rise of biometric identification (BioID) has spurred privacy concerns, leading to the creation of secure protocols for retrieval without revealing sensitive data. Though many solutions use homomorphic encryption (HE) for strong security, its computational intensity hinders real-world practicality. In this work, we consider BioID-based retrieval scenarios for decision-making purposes and propose a secure and efficient scheme, \(\textsf {SEBioID}\) . By hybridizing lightweight additive secret sharing (ASS) and garbled circuits (GC) within two-party computation (2PC), \(\textsf {SEBioID}\) employs unique strategies and carefully designed protocols that achieve significant efficiency gains over HE-based solutions while maintaining security. Specifically, our design principles give birth to a series of novel 2PC supporting protocols, including vector-matrix multiplication, oblivious shuffling, and similarity selection, that directly operate over shares (possibly with the help of customized GC). Additionally, we adopt adaptive selection strategies based on different retrieval settings. Extensive experiments demonstrate that the supporting protocols achieve distance calculation and shuffling \(1.4{\times } \sim 11.1{\times }\) and \(1.3{\times } \sim 100{\times }\) faster, respectively, while reducing comparison GC costs by 1/3 in both time and communication, compared to state-of-the-art semi-honest schemes. In particular, \(\textsf {SEBioID}\) performs secure face recognition in a database of 128 individuals in about 1 s.