Markov temporal anomaly detection model based on deep autoencoder feature spaces

Abstract

Anomaly detection in high-dimensional temporal data remains a fundamental challenge in Internet of Things (IoT) security and low-altitude surveillance systems. While deep autoencoders excel at extracting compact feature representations, they lack explicit temporal modeling capabilities. Existing approaches typically rely on recurrent neural networks for sequence modeling, which may introduce additional computational overhead and reduce interpretability. This paper proposes a Markov Temporal Anomaly Detection framework based on Deep Autoencoder Feature Space (DA-MTM), which combines deep representation learning with stochastic temporal modeling. A stacked denoising autoencoder is employed to compress high-dimensional input into compact latent representations, followed by clustering to transform continuous feature spaces into discrete states. A first-order Markov transition model is then constructed to capture normal temporal state evolution, and an anomaly score is derived by jointly considering reconstruction error and transition probability. The proposed framework provides an interpretable detection mechanism by modeling temporal dependencies in the learned feature space while maintaining low computational cost. Experiments on the UNSW-NB15 dataset demonstrate that DA-MTM achieves competitive detection performance with an AUC-ROC of 0.987 and an F1-score of 0.964 alongside sub-millisecond inference latency. Although reconstruction-based detection achieves slightly higher accuracy on this benchmark, DA-MTM provides additional interpretability through explicit temporal transition modeling.