Closing the Gap Between Real and Anonymized Data for Training Face Detection Models

Data protection legislation, such as the European Union's General Data Protection Regulation (GDPR), poses significant challenges for companies when processing, storing and sharing image data. Although straightforward anonymization techniques, like face pixelization, can satisfy compliance requirements, they significantly distort facial details. However, these are critical for training and evaluating machine learning models designed to process facial features. To overcome this limitation, recent research has explored synthesis-based privacy enhancing techniques. These methods provide a promising balance between privacy protection and data utility retention. However, models trained on anonymized datasets typically still underperform compared to those trained on original data. In this study, we investigate the performance gap for face detection models trained on data anonymized using StablePrivacy, a state-of-the-art synthesis-based privacy enhancement method. Through systematic analysis, we pinpoint two key factors that degrade the data utility of the anonymized data: (1) insufficient quality of the synthetic source faces guiding the anonymization process and (2) inpainting artifacts that can lead to overfitting in face detection models. Finally, we propose mitigation strategies to address these issues while preserving robust privacy protection, thereby narrowing the performance gap for face detection tasks.