Multi-Template Siamese Networks with Clique Contrastive Learning for Palmprint Authentication

Palmprint authentication systems are rapidly gaining significant attention as a secure and user-friendly biometric modality, especially suited for mobile applications. These systems leverage palm features that are stable, unique, and easily captured with standard mobile camera technology. Several works have developed biometric authentication approaches, leveraging Siamese networks. However, these works show several drawbacks such as using a single ground truth of an individual (also known as a template) and cannot process multiple templates simultaneously. In this paper, we propose a novel Clique Contrastive Learning (CCL) for multi-template Siamese networks, designed to enhance palmprint authentication by effectively utilizing multiple registered templates per individual. CCL adapts the Siamese network architecture to accommodate a clique of templates, applying a specialized loss function that minimizes intra-clique distances while maximizing inter-clique distances. This allows for a more robust authentication process that effectively handles the natural variations in palmprint images. We evaluate our approach using the 11k Hands dataset, focusing on authentication performance under conditions with and without accessories. Our results demonstrate that the multi-template CCL approach consistently outperforms traditional single-template methods, demonstrating the efficacy of our approach in improving the accuracy and reliability of palmprint authentication systems.