GRAPH NEURAL NETWORKS AS A TOOL FOR MONITORING AND DETECTING ANOMALIES IN CLOUD SYSTEMS
Cloud infrastructure anomalies cause significant downtime and financial losses. Traditional anomaly detection methods fail to capture complex dependencies in microservice architectures. This paper presents a novel Temporal-Attentive Graph Autoencoder (TAGAE) framework for cloud anomaly detection, leveraging Graph Neural Networks (GNNs) to model topological relationships and temporal dynamics. Our method integrates multi-source telemetry (logs, metrics, traces) into a unified graph structure, utilizes anomaly amplification layers for enhanced sensitivity, and employs focal loss for data imbalance mitigation. Evaluated on Azure-DIAD and GCP datasets, TAGAE achieves 94.2% F1-score and 96.5% AUC-PR, reducing detection latency by 63% compared to GraphSAGE. We further analyze robustness under 40% noise/missing data and propose federated GNNs for privacy-preserving deployment.
Сведения
- Опубликовано
- 2026 · № 2 (70), С. 266-277
- Язык
- Английский
- ISSN
- 2309-6977 · 2958-003X · 2958-0048
- Издатель
- Non-profit limited company "Manash Kozybayev North Kazakhstan University
- Аффилиации
- Faculty of Computer Science and Information Technolog Universiti Putra Malaysia