Face analysis and deepfake detection

This thesis concerns deep-learning-based face-related research topics. We explore how to improve the performance of several face systems when confronting challenging variations. In Chapter 1, we provide an introduction and background information on the theme, and we list the main research questions of this dissertation. In Chapter 2, we provide a synthetic face data generator with fully controlled variations and proposed a detailed experimental comparison of main characteristics that influence face detection performance. The result shows that our synthetic dataset could complement face detectors to become more robust against specific features in the real world. Our analysis also reveals that a variety of data augmentation is necessary to address differences in performance. In Chapter 3, we propose an age estimation method for handling large pose variations for unconstrained face images. A Wasserstein-based GAN model is used to complete the full uv texture presentation. The proposed AgeGAN method simultaneously learns to capture the facial uv texture map and age characteristics.
In Chapter 4, we propose a maximum mean discrepancy (MMD) based cross-domain face forgery detection. The center and triplet losses are also incorporated to ensure that the learned features are shared by multiple domains and provide better generalization abilities to unseen deep fake samples. In Chapter 5, we introduce an end-to-end framework to predict ages from face videos. Clustering based transfer learning is used to provide proper prediction for imbalanced datasets.