Aligning Normal Representations in Diffusion Model for Video Anomaly Detection

Recent advances have highlighted the potential of diffusion models in Video Anomaly Detection (VAD). Diffusion models are typically employed to generate negative instances to distinguish them from positive ones. However, the existing diffusion model architectures, generally based on the reconstruction of low-level noisy features, introduce spurious correlations due to shortcut learning, which undermines the robustness of anomaly detection. In this work, we leverage normal-specific representations to guide behavior restoration by aligning disentangled task-relevant representations within the diffusion model. we propose a normal representation-guided conditional diffusion model for unsupervised VAD by aligning normal-specific representations. Inspired by prior knowledge of anomaly discrimination, we decompose normal behavior features into normal-specific and VAD-irrelevant representations into independent channels based on contrastive learning. We introduce a group-supervised learning strategy in learning patch-wise generation guided by normal-specific representations. A gradient-based representation alignment loss enforces the alignment between normal semantics and the target patches. This process enables the diffusion model to understand normal patterns for anomaly detection. Extensive experimental results conducted on VAD benchmarks demonstrate the effectiveness of our methods.