PD-GAN: Palmprint De-Identification Based on Generative Adversarial Networks

User-friendly palmprint recognition enhances the user experience and promotes widespread adoption of palmprint-based payment systems accross diverse consumer electronics. However, palmprints contain biometric identifiers that raise significant privacy concerns regarding identity protection. Once these original palmprint data are leaked, it will be misused by adversaries for malicious activities. Despite this growing concern, research on methods to obscure or anonymize palmprints remains largely unexplored. Therefore, this study proposes a novel approach for palmprint de-identification using generative adversarial networks, termed PD-GAN, to effectively generate de-identified palmprints and remove identifiable identity features from palmprint images while maintaining their utility, thereby reducing the risks of privacy leakage and enabling secure data sharing and utilization. To ensure the stability and controllability of the synthesis process, we innovatively introduced the SSIM loss function and a palmprint recognition loss function based on ResNet18. Additionally, we proposed three evaluation metrics including de-identification rate, SSIM, and PSNR, to assess the de-identification effectiveness of the generated palmprints and the quality of the synthesized images. On the PolyU dataset, the recognition accuracy of the palmprint images generated by PD-GAN exhibited a substantial decline while preserving usability of the images. Notably, the accuracy in the ResNet18 recognition model dropped sharply to 2%. Similarly, on the Tongji palmprint dataset, the recognition accuracy dropped to 14.06% under the same ResNet18 architecture. The experimental results demonstrate that the proposed PD-GAN effectively generates de-identified palmprints images to protect identity privacy while retaining sufficient discriminative features for new recognition tasks. This ensures the utility of de-identified palmprints in authentication systems without compromising privacy, offering a practical solution for secure palmprint-based applications.