Anomaly Detection in Microservice Environments via Conditional Multiscale GANs and Adaptive Temporal Autoencoders

Abstract

This paper proposes a microservice anomaly detection method based on the fusion of Generative Adversarial Networks and temporal autoencoders. It aims to address the problems of scarce anomalous data and insufficient detection accuracy in distributed systems. The proposed framework consists of two core modules: a Conditional Multi-Scale Feature-enhanced Generative Adversarial Network (CMSF-GAN) and an Adaptive Threshold Temporal Autoencoder (ATTAE). CMSF-GAN generates diverse and semantically consistent anomalous traffic samples by using prior knowledge of anomaly types and a multi-scale feature extraction mechanism. This improves the anomaly coverage during the training phase. ATTAE models multivariate time series data through an LSTM structure. It introduces a dynamic threshold adjustment mechanism to achieve high sensitivity in detecting complex and subtle anomalies. Extensive experiments are conducted on two public datasets, Alibaba Cluster Trace 2018 and SWaT. The results are compared with several state-of-the-art methods. The proposed model demonstrates advantages in accuracy, generalization, and robustness. In addition, transferability evaluation, perturbation intensity tests, and hyperparameter sensitivity analysis further show the model's stability and practical potential in complex scenarios.