Towards adversarial realism and robust learning for IoT intrusion detection and classification

The Internet of Things (IoT) faces tremendous security challenges. Machine learning models can be used to tackle the growing number cyber-attack variations targeting IoT systems, but increasing threat posed by adversarial attacks restates need for reliable defense strategies. This work describes types constraints required a realistic example and proposes methodology trustworthy robustness analysis with evasion attack vector. proposed was evaluate three supervised algorithms, Random Forest (RF), Extreme Gradient Boosting (XGB), Light (LGBM), one unsupervised algorithm, Isolation (IFOR). Constrained examples were generated Adaptative Perturbation Pattern Method (A2PM), performed against created regular training. Even though RF least affected in binary classification, XGB consistently achieved highest accuracy multi-class classification. obtained results evidence inherent susceptibility tree-based algorithms ensembles demonstrates benefits training design approach more robust network intrusion detection