SynPalms: Palm Identification with Synthetic Data

Abstract

Person identification systems based on biometric modalities, such as hand images, require extensive and diverse datasets for effective training. However, real-world datasets are often limited in size and variability, leading to poor generalization, while strict privacy regulations constrain their sharing and use. To address these issues, we introduce the SynPalms framework, an approach for generating diverse palmar-side hand images of novel synthetic identities achieved with identity mixing during sampling. We investigate the impact of synthetic data on a ResNet50-based identification system by comparing the classification accuracy obtained when training with either real-world or synthetic images, as well as when utilizing both real and synthetic data. Our experiments demonstrate the trade-off between classification accuracy and subject privacy, highlighting the potential of synthetic data for training biometric identification systems in a privacy-preserving manner.