Face deidentification with generative neural networksMeden, Blaž (Avtor) Peer, Peter (Mentor) Štruc, Vitomir (Komentor) face deidentificationgenerative neural networksfacial synthesisprivacy protectiondata utilityimage-based biometricsAdvancements in image-based biometrics and the increasing reliance on facial data have raised concerns regarding privacy protection and the potential misuse of personal information. Face deidentification techniques have emerged as a promising approach to mitigate privacy risks while preserving data utility. This thesis investigates the application of generative neural networks for facial synthesis to achieve effective face deidentification while maintaining the usefulness of the data for subsequent biometric analysis. The primary objective of this research is to develop novel techniques for face deidentification using generative neural networks. By leveraging generative deep learning algorithms, realistic synthetic faces are generated, which substitute the source facial features while preserving essential non-identity-related characteristics. Privacy protection is a critical aspect of this research, with a focus on deidentifying facial images to prevent unauthorized identification of individuals. Various techniques such as face swapping, the utilization of formal privacy mechanisms, preservation of facial attributes, and identity suppression are explored to ensure that the synthesized faces remain untraceable while maintaining their realistic appearance and data utility for further analysis. Furthermore, the thesis addresses the challenges of evaluating the effectiveness of face deidentification techniques in terms of both privacy protection and data utility. Metrics and benchmarks are presented to quantify the level of anonymity achieved while measuring the impact on data utility through the analysis of preserved facial attributes. The evaluation process involves comparing recognition accuracy, facial attribute classification performance, and other image quality metrics on deidentified face images against the source facial images. The findings of this thesis contribute to advancing face deidentification and privacy protection of biometric data, providing competitive practical solutions for face deidentification by utilizing state-of-the-art generative models.20232023-11-21 11:40:08Doktorsko delo/naloga152341sl