FAU-Preserving Video Anonymization for Privacy-Aware Reproducible Research

Abstract

Video datasets containing identifiable faces cannot be legally shared, creating a significant barrier to privacy-aware reproducible research in affective computing and HCI. Common anonymization methods such as blurring and pixelation destroy Facial Action Units (FAUs), the facial expression signals that make behavioral video data scientifically useful. This research investigates how much FAU information survives anonymization and under what conditions, treating it as a privacy–utility measurement problem. I propose an evaluation framework that quantifies identity removal via ArcFace cosine similarity and FAU preservation via temporal Pearson correlation with OpenFace 3.0. To test this framework, I build a self-supervised pipeline trained directly on real unlabelled video, without synthetic data. The pipeline uses RetinaFace (MobileNetV2, uniface) for detection, face-distance-adaptive pixelation, and Conditional Flow Matching (CFM) for expression-consistent reconstruction. The CFM U-Net is trained with a composite loss balancing reconstruction quality, FAU preservation, and identity removal. IoU tracking with noise seed locking and EMA smoothing maintains temporal coherence across frames.