PRISM: Privacy-Aware Federated Periocular Identification under Extreme Label Skew for Mobile Devices

Periocular biometrics have demonstrated strong potential for mobile authentication. However, training accurate recognition models typically requires centralising biometric data, which conflicts with privacy regulations such as EU GDPR and AI Act. This study proposes PRISM, a privacy-aware federated learning approach for periocular identification that enables collaborative model training across devices without sharing raw biometric data. PRISM tackles the realistic one-identity-per-device setting of mobile biometrics, enabling federated learning under extreme label skew while keeping all biometric templates strictly on-device. A key component is Classifier Prototype Synchronisation (CPS), a novel mechanism that reconstructs identity-specific classifier prototypes from server-side weights after aggregation, avoiding the need to transmit biometric embeddings. Evaluated on the popular UFPR-Periocular database (1,122 identities), PRISM improves Rank-1 accuracy by 5.9 pp over standard federated learning, closing by 43% the performance gap with respect to centralised training. To the best of our knowledge, this is the first study applying federated learning to periocular identification under realistic mobile deployment constraints.