Emotion-Aware Face De-identification with Generative Adversarial Networks

Abstract This paper presents an innovative end-to-end emotion-preserving de-identification approach utilizing a Generative Adversarial Network (GAN), employing the StyleGAN architecture. The method produces natural-looking de-identified images by generating a synthetic face image dataset and utilizing the DeepFace model for gender classification and representative image selection. An enhanced SimSwap framework enhances emotion preservation, accompanied by a novel loss function tailored to emotional expression preservation. The deep face model serves to classify and recognize emotional expressions in both original and de-identified swapped images. Explicit preservation of emotional expressions during face-swapping is achieved through attribute preservation loss minimization. The paper conducts a detailed ablation study to demonstrate the superiority of the proposed GAN components. The method exhibits superior performance in accuracy and emotion preservation compared to recent face de-identification methods.