Learning Multi-Scale Generative Representations for Cloud Performance Anomaly Detection via Self-Distillation

Abstract

This paper addresses the challenges of performance anomaly detection in cloud service systems under highly dynamic topologies, complex dependency structures, and multi-source heterogeneous data conditions, and proposes a detection framework based on a self-distillation multi-scale diffusion model. The proposed method employs a multi-scale diffusion mechanism to model features at different temporal granularities, capturing both temporal evolution and spatial dependencies of system performance during the generation and reconstruction processes, thereby forming globally consistent and locally sensitive representations in the feature space. The model consists of four core components: multi-scale encoding, diffusion generation, self-distillation constraint, and anomaly determination, where the diffusion process models dynamic changes in performance distributions and the self-distillation mechanism enhances cross-scale feature consistency and stability. Using multidimensional monitoring data from a cloud platform, the study conducts comparative and sensitivity experiments to evaluate the effects of learning rate, diffusion steps, time window length, and anomaly ratio on detection performance. Experimental results show that the proposed model outperforms mainstream methods in terms of accuracy, precision, recall, and F1 score, achieving efficient anomaly identification and feature aggregation under unsupervised conditions, and demonstrating strong robustness and generalization ability. The findings confirm the effectiveness of the multi-scale diffusion and self-distillation mechanisms for performance anomaly detection in complex cloud environments, providing a new generative modeling solution for intelligent cloud operations and system stability assurance.