Selective Noise Injection and Feature Scoring for Unsupervised Request Anomaly Detection

Abstract

This paper addresses the problem of anomaly detection in high-dimensional and complex request scenarios. An unsupervised anomaly request detection method based on diffusion models is proposed. The method integrates generative modeling with a feature discrimination mechanism. It builds a forward noise injection and reverse reconstruction process to effectively capture the distribution characteristics of normal request data. Specifically, a structure-aware module, Selective Noise Injection for Reconstruction (SNIR), is introduced. It selectively injects noise during the forward diffusion phase to preserve key feature dimensions and improve reconstruction quality. On this basis, a Feature-aware Discriminative Scoring (FDS) mechanism is designed. It embeds the semantic features of original and reconstructed requests and computes a combined score using Euclidean distance and cosine similarity. This enables fine-grained discrimination of abnormal requests. The method does not rely on labeled data. It uses only normal samples for both modeling and detection. It offers strong generalization and practical applicability. Experimental results on multiple benchmark datasets show that the proposed method significantly outperforms existing representative methods in terms of AUC, Precision, and F1. It effectively handles diverse attack patterns and changes in data distribution. It also maintains stable detection performance in high-noise environments.