Anomaly Detection and Localization of IoT Traffic via Hybrid Autoencoder and Temporal Modeling

To address the problems of inadequate anomaly representation in IoT traffic and poor anomaly localization capabilities, this study proposes a novel anomaly detection and localization framework based on probabilistic autoencoder reconstruction, temporal alignment, and multi-scale temporal feature learning. The model uses a Variational Autoencoder to learn the distribution of normal traffic and performs sequence reconstruction. It combines Dynamic Time Warping to improve the identification of time-misaligned anomalies such as rhythm offset and phase drift. It then feeds anomalous windows into a Transformer-based encoding structure with inductive bias to model global temporal correlations. It uses Multi-Scale Feature Fusion decoding to recover local details step by step, and achieves joint localization of anomalous time slices, key variables, and related nodes. Experimental results indicate that with 28,800 test cases, the model achieves 98.74% of the area under the curve, 95.96% of the F1 score at the window level, 2.4 s of temporal localization errors, and 92.68% of node backtracking accuracy. In practical traffic edge environments, the recall rate of typical anomaly detection reaches up to 97.83%, with a response delay of 54 ms and an early warning lead time of 55 s. These results show that the model achieves high-accuracy detection and fine-grained localization in complex edge traffic environments, and provides support for network security monitoring and anomaly tracing.