De-identification of profile pictures while retaining facial features using generative adversarial networks

• 主要作者: Lim, Yee Han
• 其他作者: Kong Wai-Kin Adams
• 格式: Final Year Project
• 語言: English
• 出版: Nanyang Technological University 2021
• 主題: Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
• 在線閱讀: https://hdl.handle.net/10356/148060

In this paper, we propose a method to perform de-identification of profile pictures by using Generative Adversarial Networks (GANs) to generate similar images of people. The goal is to maximally mask the identity of the individual, by making them unrecognisable to their friends and family, while retaining important facial features that make them look indistinguishable to strangers. We dub this goal, “Unrecognisable To Friends, Indistinguishable To Strangers” (UTF-ITS). We achieve this by minimising the L2 loss between the reference image and the generated image, using a pre-trained model published in NVIDIA’s StyleGAN research. By prematurely stopping the minimisation of L2 loss at a desirable iteration, we can achieve the UTF-ITS goal. This process could be useful for social applications where it is necessary to reveal an individual’s facial likeness, but users might not want to reveal their identity for privacy reasons, i.e. Dating Apps, Online Forums.