Privacy-Aware Video Anomaly Detection through Orthogonal Subspace Projection

Video anomaly detection (VAD) systems often prioritize accuracy while overlooking privacy concerns, limiting their suitability for real-world deployment. We propose the Orthogonal Projection Layer (OPL), a lightweight module that removes task-irrelevant variations to produce representations focused on anomaly-relevant cues. To address privacy risks in human-centered scenarios, we introduce Guided OPL (G-OPL), which suppresses facial attributes using weak supervision from face-presence signals while preserving non-identifying features such as pose and motion. A cosine alignment objective enforces consistent capture and removal of facial information without identity labels or adversarial training. We further present a privacy-aware evaluation framework that jointly assesses detection performance and privacy preservation, and enables analysis of how sensitive information is filtered. Experiments show that embedding privacy constraints into model design reduces sensitive information while maintaining or improving detection accuracy, supporting projection-based architectures as a principled approach for privacy-aware VAD.

Modern video surveillance systems can automatically detect unusual or dangerous events, helping improve safety in places such as airports, train stations, and shopping centers. However, these systems often collect sensitive personal information, especially people’s faces, even when this information is not necessary for detecting abnormal events. This creates important privacy concerns and may reduce public trust in AI technologies. In this paper, we develop a new AI approach that helps video analysis systems focus on important events while reducing the amount of private information stored in their internal representations. Our method removes unnecessary and privacy-sensitive information, such as facial identity cues, while preserving useful information related to human actions and motion. Unlike many existing privacy-preserving approaches, our method is lightweight, easy to integrate into existing systems, and does not require identity labels or complicated training procedures. We also introduce new ways to measure both anomaly detection performance and privacy protection together. Experiments on multiple benchmark datasets show that our method can significantly reduce privacy leakage while maintaining or even improving detection accuracy. Our work contributes toward building safer, more trustworthy, and privacy-aware AI systems for real-world video monitoring.