Discriminative Reconstruction Method for Backend Anomaly Detection Based on Diffusion-Generated Prior Enhancement

Authors
  • Jingyi Chen

    Northeastern University, Silicon Valley, USA

    Author

Abstract

To address the challenges of concealed anomaly patterns, enhanced operational state coupling, and heterogeneous dispersion of multi-source monitoring information in complex backend systems, this paper proposes a discriminative reconstruction backend anomaly detection method based on diffusion-generated prior enhancement. This method utilizes backend operational information such as logs, metrics, and call chains. First, it constructs a unified latent representation through feature encoding to alleviate semantic fragmentation caused by heterogeneous observations. Then, a diffusion-generated mechanism is introduced to learn the latent priors of normal operating states, enabling the model to achieve more stable structural awareness under noise disturbances and complex dependencies. Based on this, the enhanced prior representation is fed into the discriminative reconstruction branch. Through reconstruction consistency constraints and discriminative modeling, the deviation of anomaly samples from normal patterns is jointly characterized. Furthermore, a joint anomaly scoring method is designed to achieve comprehensive identification of complex backend anomalies. The constructed method simultaneously focuses on latent structural recovery, anomaly semantic shift, and discriminative boundary enhancement, thereby improving the identification ability of concealed anomalies, weak anomalies, and anomalies with ambiguous boundaries. Comparative experimental results show that this method has good overall performance and application value in backend anomaly detection tasks, and can provide effective support for anomaly perception and intelligent operation and maintenance analysis in complex distributed systems.

Cover Image
Downloads
Published
2024-06-01
Section
Articles