Facial image super-resolution network for confusing arbitrary gender classifiers

Existing facial image super-resolution methods have identified the capacity to transform low-resolution facial images into high-resolution ones. However, clearer high-resolution facial images increase the possibility of accurately extracting soft biometric features, such as gender, posing a significant risk of privacy leakage. To address this issue, we propose a gender-protected face super-resolution network, which can incorporate gender-identified privacy information by introducing fine image distortion during the super-resolution process. It progressively transforms low-resolution images into high-resolution ones while partially disturbing the face images. This procedure ensures that the generated super-resolution facial images can still be utilized by face matchers for matching purposes, but are less reliable for attribute classifiers that attempt to extract gender features. Furthermore, we introduce leaping adversarial learning to help the super-resolution network to generate gender-protected facial images and work on arbitrary gender classifiers. Extensive experiments have been conducted using multiple face matchers and gender classifiers to evaluate the effectiveness of the proposed network. The results also demonstrate that our proposed image super-resolution network is adaptable to arbitrary attribute classifiers for protecting gender privacy, while preserving facial image quality. • Presenting a facial super-resolution method with gender privacy protection. • Designing a GAN-based model to achieve gender-protected facial super-resolution. • Leveraging leaping adversarial training to generalize the super-resolved facial images.