Abstract
Cloud infrastructure systems have become increasingly critical for modern enterprises, yet their complexity poses significant challenges for anomaly detection, particularly when labeled training data is scarce. This paper proposes a novel framework that integrates self-supervised contrastive learning with few-shot learning techniques to address the challenge of anomaly detection in cloud environments with limited labeled samples. The approach leverages contrastive representation learning to extract meaningful features from unlabeled cloud system data, enabling the model to effectively distinguish between normal and anomalous patterns using only a few labeled examples per anomaly class. Our method introduces a multi-stage architecture that combines variational autoencoder-based representation learning with prototypical network classification and meta-learning optimization strategies. Experimental results on three real-world cloud infrastructure datasets demonstrate that the proposed framework achieves superior performance compared to traditional supervised methods and existing few-shot learning approaches, with an average F1-score improvement of 12.3% in 5-way 1-shot scenarios. The framework exhibits strong generalization capabilities across different cloud service types and infrastructure configurations, making it particularly suitable for practical deployment in dynamic cloud environments where new anomaly patterns emerge frequently.

This work is licensed under a Creative Commons Attribution 4.0 International License.
Copyright (c) 2025 Liwen Cheng, Haoqi Sun, Elena Petrović (Author)