@article{Geng2026, 
author = {Haijun Geng and Zhi Zhang and Yuhua Qian and Bo Yang and Qi Ma and Haotian Chi and Jing Yang and Xia Yin},
title = {CAE-IF: An Anomaly Detection Approach Based on Temporal Representation of the Reconstruction Error},
year = {2026},
journal = {Tsinghua Science and Technology},
volume = {31},
number = {4},
pages = {2055-2070},
keywords = {unsupervised learning, network traffic detection, reconstruction error, temporal representation},
url = {https://www.sciopen.com/article/10.26599/TST.2024.9010215},
doi = {10.26599/TST.2024.9010215},
abstract = {In the field of network traffic anomaly detection, unsupervised learning plays a critical role yet encounters significant challenges, including accurately determining anomaly thresholds and modeling the intricate temporal dynamics of network traffic. To address these challenges, we present a novel approach, termed Convolutional Autoencoder-Isolation Forest (CAE-IF). By leveraging packet-level reconstruction errors with contextual information, our approach obviates the need for manual threshold setting and effectively captures temporal dynamics. The process commences with the application of the damped incremental statistics algorithm to extract statistical features from network traffic with temporal information. Subsequently, the Convolutional Autoencoder (CAE) is employed to compute the Root Mean Square Error (RMSE), offering detailed insights into the temporal correlations in network traffic. This RMSE is then refined through an aggregation mechanism based on source IP addresses, yielding a fine-grained temporal representation. Finally, the Isolation Forest (IF) algorithm is applied to establish an anomaly detection framework. Our comprehensive experimental evaluation, using three datasets: Mirai, Operating System Scan (OS Scan), and Simple Service Discovery Protocol (SSDP) Flood, demonstrates the superior efficacy of the CAE-IF method. It achieves remarkable  F1 scores of  96.14%,  99.81%, and  99.98% on these datasets, respectively. These results not only signify substantial improvements over existing methods for the Mirai and OS Scan datasets but also match the highest  F1 score obtained on the SSDP Flood dataset.}
}