ИССЛЕДОВАНИЕ УСТОЙЧИВОСТИ АНСАМБЛЕВЫХ МОДЕЛЕЙ МАШИННОГО ОБУЧЕНИЯ К СОСТЯЗАТЕЛЬНЫМ АТАКАМ В СИСТЕМАХ ОБНАРУЖЕНИЯ ВТОРЖЕНИЙ IOT-СЕТЕЙ

Abstract

This paper investigates the robustness of ensemble machine learning models against adversarial attacks in the context of multiclass intrusion detection in IoT networks. The experimental framework is built upon the ToN-IoT dataset, widely used for evaluating intrusion detection systems. Three ensemble models are employed for multiclass network traffic classification: XGBoost, Random Forest, and LightGBM. Data preprocessing includes removal of irrelevant features, one-hot encoding of categorical attributes, standardization, and train-test splitting. Principal Component Analysis (PCA) is applied to examine the effect of dimensionality reduction on model robustness. To evaluate model robustness, two types of attacks are implemented: Gaussian Noise Attack and FGSM (Fast Gradient Sign Method). Since ensemble models are non-differentiable, adversarial examples for FGSM are generated via a surrogate neural network using a transfer-based approach. Additionally, adversarial training is investigated as a defense strategy, involving augmentation of the training set with adversarial examples followed by model retraining. Classification quality is assessed using Accuracy and F1-score metrics, with analysis of performance degradation under varying noise levels and perturbation magnitudes. Model behavior is interpreted through SHAP analysis. The results demonstrate high accuracy across all three models on clean data, with a notable decline in performance under adversarial attacks. Random Forest exhibits the greatest robustness among the evaluated methods. It is established that applying PCA improves model robustness, while adversarial training effectively mitigates the impact of FGSM attacks. The findings confirm the promise of ensemble methods combined with adversarial training for building reliable IDS in IoT environments.