Privacy-Aware Video Anomaly Detection through Orthogonal Subspace Projection

Lei Wang, Wenxiang Diao, Andrew Busch, Jun Zhou, Yongsheng Gao

Keywords: anomaly detection, privacy-preserving, interpretable AI, learning representations, model architecture

Abstract: Video anomaly detection (VAD) is central to modern surveillance, yet most existing methods optimize for accuracy while overlooking critical ethical concerns such as privacy and transparency. For deployment in real-world settings, VAD should not only detect anomalies reliably but also respect fundamental privacy principles. We propose the Orthogonal Projection Layer (OPL), a lightweight architectural module that suppresses task-irrelevant variations, including background clutter and noise, to produce representations focused on anomaly-relevant cues. Faces, unlike other cues such as gait or body pose, are highly sensitive biometric identifiers: they uniquely reveal identity, are tightly regulated by data protection laws, and pose immediate risks of misuse. To address the privacy risks inherent in human-centered anomalies, we extend this idea to the Guided OPL (G-OPL). Using only weak supervision from face-presence indicators, G-OPL selectively removes facial attributes while retaining non-identifying human features needed for anomaly detection. A cosine alignment loss ensures that facial information is systematically captured and neutralized, without requiring identity labels or adversarial training. We further introduce a privacy-aware evaluation framework that jointly assesses anomaly detection accuracy, privacy preservation, and interpretability. Our analysis uncovers how projection layers filter sensitive information, why this improves transparency, and under what conditions ethical design also enhances robustness. Extensive experiments confirm that embedding ethical constraints directly into model design strengthens privacy protection while maintaining, and in some cases improving, anomaly detection performance. These results position projection-based architectures as a principled path toward trustworthy and deployable VAD systems.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 6588