Learning Structural Similarity with Evolutionary-GAN: A New Face De-identification Method

With the explosive growth of image sources, the development of face recognition technology and the emphasis on privacy, face deidentification has become increasingly important. The purpose of face de-identification is to hide the identity of individuals in a video or image while still retaining certain facial attributes. The mainstream methods of face de-identification are mostly based on the k-same framework, which generates anonymized faces that are not diverse and have poor visual quality. This paper presents a face de-identification method that uses an improved evolutionary generative adversarial network to synthesize faces to de-identificate, using the structural similarity index and the distance between the original face and the de-identificated face for generator selection, choosing the optimal generator to enter the next round of evolution. We have proved the feasibility of the proposed method through extensive experiments, indicating that our alteration method is indeed effective.