Privacy-Aware Video Anomaly Detection via Orthogonal Subspace Projection

Researchers have introduced a novel approach to Video Anomaly Detection (VAD) that addresses critical privacy concerns often overlooked in favor of accuracy. The study proposes the Orthogonal Projection Layer (OPL), a lightweight module designed to eliminate task-irrelevant variations, focusing representations on anomaly-relevant cues. To specifically mitigate privacy risks in human-centric scenarios, the authors developed Guided OPL (G-OPL). This mechanism suppresses facial attributes using weak supervision from face-presence signals while preserving non-identifying features like pose and motion. A cosine alignment objective ensures consistent removal of facial information without requiring identity labels or adversarial training. Additionally, the team presents a privacy-aware evaluation framework that jointly assesses detection performance and privacy preservation. Experimental results demonstrate that embedding these privacy constraints into the model design effectively reduces sensitive information exposure while maintaining or even improving detection accuracy. This work supports projection-based architectures as a principled solution for deploying privacy-aware VAD systems in real-world applications, balancing security needs with individual privacy rights.