Unsupervised Anomaly Detection in Structured Data Using Structure-Aware Diffusion Mechanisms

Abstract

To address the challenge of limited accuracy in detecting anomalies within complex structured data, this paper proposes a structure-aware anomaly detection method based on diffusion models. The method builds a generative diffusion framework that models the distribution of normal data through a forward noise perturbation and reverse denoising process. This enables the identification of abnormal data. To enhance the model's understanding of dependencies among fields in structured data, a Structure-Aware Diffusion (SAD) mechanism is introduced. It uses a structural matrix to explicitly model the semantic and logical relationships between fields, allowing the diffusion process to follow structural constraints. In addition, a Dynamic Reconstruction Scoring (DRS) mechanism is proposed. During the anomaly scoring phase, it dynamically adjusts weights based on the reconstruction uncertainty of different fields. This improves detection accuracy for local and sparse anomalies. Experiments on public datasets show that the proposed method outperforms traditional neural network models and baseline generative models in terms of Accuracy, AUC, and F1-score. It effectively identifies multiple types of structured anomalies. Further ablation studies confirm the significant contributions of the two proposed mechanisms. The results demonstrate the effectiveness of structure awareness and dynamic scoring in high-dimensional structured anomaly detection tasks. By integrating generative learning with structural information, this paper provides a high-accuracy, generalizable anomaly detection approach for complex relational data. The method shows strong robustness and practical value.