A Contrastive GAN‐Based Framework for Full‐Body Visual Privacy Protection in Open World Scenarios

ABSTRACT Generative adversarial networks (GANs) with their strong generative capabilities have shown significant promise in visual privacy protection. However, when applied to image full‐body visual privacy protection in open‐world scenarios, where abnormal visual privacy data may exist in the training data, issues such as mode collapse and instability in GANs can be severely exacerbated. This leads to a significant reduction in both image quality and utility preservation. In this paper, we propose an end‐to‐end, contrastive GANs‐based framework, FBPPGAN, for image full‐body visual privacy protection, specifically designed to address these challenges. First, we introduce the architecture of FBPPGAN, which is tailored for full‐body visual privacy protection. Second, we propose a novel adversarial loss function aimed at mitigating mode collapse and instability, particularly in the presence of abnormal images in open‐world environments. We also design a content mapping network and a content loss function based on contrastive learning to address the issue of insufficient color gamut in generated images. Furthermore, a stylized loss function is introduced to more accurately measure the difference between the generated and target domains. Experimental results across four public datasets demonstrate that FBPPGAN effectively overcomes mode collapse and instability, delivering superior image quality and utility preservation. Compared to the existing state‐of‐the‐art methods, FBPPGAN outperforms in terms of convergence, stability, computational complexity, processing speed, and effectiveness. To the best of our knowledge, this is the first GAN‐based framework for image full‐body visual privacy protection in open‐world scenarios.