Characterizing Robustness of Tabular Intrusion Detection System Models Under Sensor Degradation: A Systematic Empirical Evaluation for Unmanned Aerial Vehicle Ad Hoc Networks
Concurrency and Computation: Practice and Experience · 2026-11 · 38 (22)
DOI: 10.1002/cpe.71022 ↗Kurz gefasst
Fachartikel in der Zeitschrift Concurrency and Computation: Practice and Experience, online erschienen am 08.11.2026. Verfasst von Qianli Di und 3 weiteren Autor:innen. Die Arbeit nennt 45 Literaturangaben.
Abstract (Originalsprache)
Intrusion detection systems (IDS) for unmanned aerial vehicle (UAV) ad hoc networks are typically evaluated under clean, complete sensor data. However, real‐world UAV deployments routinely face sensor failures, communication dropouts, and electromagnetic interference that degrade input quality. This paper presents, to the best of our knowledge, the first systematic empirical evaluation of how tabular deep learning architectures and data augmentation strategies interact with sensor degradation in UAV IDS. We evaluate 8 model‐augmentation combinations across 6 degradation scenarios (missing features, Gaussian noise, salt‐and‐pepper noise, and combined degradation) on a primary UAV benchmark (UAV‐DARPA) and a non‐UAV cross‐domain robustness check (CICIDS‐2017), with 5‐seed statistical validation and Holm–Bonferroni multiple‐comparison correction. Our results reveal three key findings. First, the FT‐Transformer architecture significantly outperforms MLP and tree‐based models under missing‐feature conditions, achieving 0.726 F 1 ‐macro at 80% missing features compared to 0.587 for MLP (, Cohen's ), a robustness advantage consistent with its per‐feature tokenization isolating missing features to individual tokens. Second, augmentation effectiveness is degradation‐type specific: MaskAug excels at missing features, NoisAug at sensor noise, and MixAug is the robust default for unknown degradation scenarios (combined F 1 = 0.814). Third, we present a negative finding that FT‐Transformer attention weights exhibit near‐uniform distributions on tabular IDS data (entropy 0.884–0.951), challenging the transfer of attention‐based feature importance methods from NLP/vision to tabular security domains. We provide five evidence‐backed guidelines for deploying robust IDS in degradation‐prone UAV networks.
Zitieren
APA
Di, Q., Du, K., Chen, G., & Deng, H. (2026). Characterizing Robustness of Tabular Intrusion Detection System Models Under Sensor Degradation: A Systematic Empirical Evaluation for Unmanned Aerial Vehicle Ad Hoc Networks. Concurrency and Computation: Practice and Experience, 38(22). https://doi.org/10.1002/cpe.71022
Vancouver
Di Q, Du K, Chen G, Deng H. Characterizing Robustness of Tabular Intrusion Detection System Models Under Sensor Degradation: A Systematic Empirical Evaluation for Unmanned Aerial Vehicle Ad Hoc Networks. Concurrency and Computation: Practice and Experience. 2026;38(22). doi:10.1002/cpe.71022
BibTeX
@article{Di2026,
author = {Di, Qianli and Du, Keju and Chen, Gang and Deng, Hai},
title = {Characterizing Robustness of Tabular Intrusion Detection System Models Under Sensor Degradation: A Systematic Empirical Evaluation for Unmanned Aerial Vehicle Ad Hoc Networks},
journal = {Concurrency and Computation: Practice and Experience},
year = {2026},
volume = {38},
number = {22},
doi = {10.1002/cpe.71022}
}
EUVIMED