AI-Augmented Anomaly Detection via Generative Distribution Modeling and Uncertainty Quantification in Cloud Systems

Abstract

This study proposes an anomaly recognition framework for cloud computing environments characterized by high dynamism, strong coupling, and uncertainty. The method integrates generative adversarial learning with uncertainty estimation to achieve robust detection under unsupervised conditions. By constructing an adversarial structure between the generator and the discriminator, the model learns the normal distribution of the system and detects anomalies through reconstruction errors. An uncertainty estimation mechanism is introduced to quantify prediction confidence, ensuring stable detection performance under noise interference, sampling bias, and distribution drift. The model consists of five stages: feature encoding, latent space mapping, generative reconstruction, distribution discrimination, and confidence regulation. Through joint optimization of adversarial loss and reconstruction error, the model achieves global consistency and local separability in the feature space. Sensitivity analyses are conducted across hyperparameters, environmental conditions, and data proportions, including variations in latent vector dimension, anomaly ratio, system load, and sampling number. Experimental results show that the proposed method significantly outperforms mainstream algorithms in Accuracy, Precision, Recall, and F1-Score, maintaining stable performance under high load, noise enhancement, and varying anomaly ratios. The findings demonstrate the effectiveness of combining generative distribution modeling with uncertainty quantification and provide a scalable, interpretable, and robust solution for anomaly detection and risk assessment in cloud computing systems.