Enhancing Security and Privacy in Occluded Face Recognition: A Human‐Centered GAN‐Based Approach for Masked Identities in High‐Security Environments

Abstract

Facial masks are still a big problem for regular facial recognition systems, especially in places where security is very important. This is because so many people wear them for health, cultural, or security reasons. This study presents a multi‐stage face reconstruction system utilising generative adversarial networks, aimed at restoring occluded facial regions while maintaining identity, structural accuracy and privacy protection. The suggested method uses gender categorisation, facial landmark recognition and mask segmentation to help with a landmark‐aware inpainting procedure. Separate training trajectories for male and female faces, as well as structural priors, help make reconstructions that are more accurate and consistent with their attributes. The model's main part is an encoder‐decoder generator that was trained with a composite loss function that balances perceptual quality, adversarial realism, pixel‐level precision and semantic coherence. The method takes a biometric privacy approach, rebuilding only the facial areas needed for recognition and hiding individually identifiable or unnecessary features to protect both recognition accuracy and user privacy. We built a huge matched dataset of 70,000 masked and unmasked face images from FFHQ to use for training and testing. The suggested strategy outperforms state‐of‐the‐art inpainting techniques, as shown by quantitative findings on various common metrics, such as structural similarity index (SSIM) (0.95), PSNR (33.3 dB) and identity similarity. In addition to technological contributions, our study moves forward the creation of AI systems that are ethical and open for use in sensitive areas like surveillance, border control and other areas where security and user privacy must be carefully balanced.