Implementing an Intrusion Detection System in Internet of Things Resource-Constraint Environment Using Hybrid Extreme Gradient Boost and Gated Recurrent Unit

Intrusion Detection Systems (IDS) for Internet of Things (IoT) environments increasingly rely on deep learning models to detect complex attack patterns. While recent architectures achieve high accuracy under benign conditions, their resilience against adversarial evasion attacks remains insufficiently explored. This study evaluated adversarial robustness as primary design objective rather than secondary performance attribute. A lightweight hybrid XGBoost–GRU model was proposed and compared against attention-based LSTM (AT-LSTM). Gradient-based evasion attacks that includes Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) were employed under bounded perturbation budgets to simulate realistic adversarial behaviour. Experimental results on the UNSW-NB15 dataset showed that with the FGSM attack, XGBoost–GRU model retains approximately 90% accuracy at ε = 0.10 whereas the AT-LSTM degrades more sharply to approximately 86%. Under the PGD attacks which represent stronger and more adaptive adversarial strategy, the AT-LSTM experienced severe performance collapse with adversarial accuracy dropping to approximately 43% at ε = 0.10 while the developed XGBoost–GRU model maintains adversarial accuracy above 70% under the same conditions. Also, for the Attack Success Rate (ASR) analysis, more than half of malicious samples successfully evade detection in the AT-LSTM under PGD attacks (58%) whereas the XGBoost–GRU model limits successful evasions to less than 30%. The results from the developed XGBoost–GRU model consistently maintained higher adversarial accuracy, lower attack success rates, and superior robust accuracy retention under increased adversarial evasion. The findings demonstrate that architectural simplicity can enhance detection stability under adversarial pressure for intrusion detection in IoT systems.