PAGM: Partially-Aligned Global and Marginal Multi-View Contrastive Clustering for Facial Recognition in Consumer Electronics

Facial recognition has become a cornerstone of consumer electronics, underpinning critical functions such as device security, user authentication, and privacy-sensitive applications. As these systems become increasingly embedded in daily consumer devices like smartphones, smart home assistants, and wearables, ensuring their forensic reliability and preserving user privacy have become paramount. Despite significant advancements, traditional supervised learning methods often require substantial labeled data, making unsupervised approaches more desirable for scalable and privacy-preserving deployments. How ever, existing unsupervised facial recognition methods commonly assume perfect alignment across views, a condition that is rarely achieved in practical consumer electronics environments. Moreover, conventional techniques that combine multi-view information through concatenation or weighted fusion can introduce noise and compromise system integrity. To overcome these challenges, we introduce Partially-Aligned Global and Marginal Multi-View Contrastive Clustering (PAGM), a novel method specifically designed to handle partially aligned multi-view facial data. PAGM employs both global and marginal contrastive losses to enhance discriminative feature learning while maintaining privacy and data integrity. The global learning module is inspired by the transformer architecture, employing the Query (Q), Key (K), and Value (V) matrices to construct an attention mechanism that more effectively fuses cross-modal information. Meanwhile, the margin learning module optimizes the network by treating aligned samples as positive pairs and introducing a margin term to prevent trivial solutions. Extensive experimental results validate that PAGM significantly outperforms existing methods, particularly in challenging, unaligned multi-view facial recognition scenarios in consumer electronic devices, thereby offering a promising solution for privacy-preserving biometric systems.