L-GAN: landmark-based generative adversarial network for efficient face de-identification

A large amount of high-quality data are collected through autonomous vehicles, CCTVs, guidance service robots, and web map services (Google Street View). However, the data collected through them include personal information such as peoples’ faces and vehicle license plates. Currently, personal information contained in data is de-identified using methods such as face blur, pixelation, and masking. Consequently, it loses value as data for artificial intelligence (AI) learning. Therefore, in this study, we propose a model to generate a fake face that maintains the basic structure of the human face. There are several methods for generating faces. One is to generate them using a generative adversarial network (GAN) model. The GAN is an AI algorithm used for unsupervised learning and is implemented by a system in which two neural networks compete. However, because GAN operates as an input of a random noise vector, it is difficult to obtain results for the desired face angle and shape. Therefore, pre- and post-processing is required to generate a fake face that maintains the basic structure and angles; however, the calculation is complicated, and it is difficult to generate a natural image. To solve this problem, we propose a method for generating a fake face that maintains the basic structure and angle of the real face by applying a facial landmark. Using the proposed method, it was possible to generate a fake face with a different impression while maintaining the basic structure and angle of the face.