@Article{jcs.2025.072740,
AUTHOR = {Chunyong Yin, Williams Kyei},
TITLE = {ARAE: An Adaptive Robust AutoEncoder for Network Anomaly Detection},
JOURNAL = {Journal of Cyber Security},
VOLUME = {7},
YEAR = {2025},
NUMBER = {1},
PAGES = {615--635},
URL = {http://www.techscience.com/JCS/v7n1/65064},
ISSN = {2579-0064},
ABSTRACT = {The evolving sophistication of network threats demands anomaly detection methods that are both robust and adaptive. While autoencoders excel at learning normal traffic patterns, they struggle with complex feature interactions and require manual tuning for different environments. We introduce the Adaptive Robust AutoEncoder (ARAE), a novel framework that dynamically balances reconstruction fidelity with latent space regularization through learnable loss weighting. ARAE incorporates multi-head attention to model feature dependencies and fuses multiple anomaly indicators into an adaptive scoring mechanism. Extensive evaluation on four benchmark datasets demonstrates that ARAE significantly outperforms existing autoencoder variants and classical methods, with ablation studies confirming the critical importance of its adaptive components. The framework provides a robust, self-tuning solution for modern network intrusion detection systems without requiring manual hyperparameter optimization.},
DOI = {10.32604/jcs.2025.072740}
}