Toward High Accuracy and Strong Security: Cancellable Templates for Multimodal Biometric Recognition Based on Feature Fusion

Multimodal biometric recognition integrates complementary information from different modalities and significantly improves recognition accuracy and system security compared with unimodal methods. However, with the widespread adoption of these systems, the risks of biometric template leakage and theft have emerged, posing serious threats to user privacy. To address this issue, we propose a secure multimodal template protection framework based on palmprint and palmvein. We design a feature-level fusion network, PalmSynNet, where the Multi-Scale Local Feature Extraction module enhances the extraction of local features, and the Global Feature Fusion module enables cross-modal global feature interaction, resulting in robust and highly discriminative fused representations. In addition, we propose a novel cancellable random projection method, which generates protected templates through SoftMax-based random projection (SoRP) combined with other hashing algorithms, effectively avoiding the reversibility problem of random projection under certain conditions. Extensive evaluations conducted on three multimodal palmprint databases—PolyU, TJ, and CUMT—demonstrate that proposed framework not only achieves superior fusion recognition performance compared with state-of-the-art methods, but also satisfies essential security requirements, including irreversibility, revocability, unlinkability, and resistance to various attacks.