The increasing integration of Industrial Control Systems (ICS) with IT networks has increased their vulnerability to sophisticated cyber threats, which makes the implementation of robust Intrusion Detection Systems (IDS) essential. Although Machine Learning (ML) improves detection performance, these models remain inherently vulnerable to adversarial attacks, where input data is deliberately manipulated to cause misclassification. This study investigates the robustness of ML-based IDS against such adversarial threats in an ICS environment. The study begins by addressing class imbalance in the ICS dataset and establishes that Random Forest combined with Synthetic Minority Over-sampling Technique (SMOTE) delivers the best performance according to cross-validation results. The model’s effectiveness was subsequently evaluated across different class distributions to establish a reliable baseline. A critical grey-box attack scenario was simulated using the Jacobian-based Saliency Map Attack (JSMA) and Fast Gradient Sign Method (FGSM). Results demonstrated a notable decline in model performance, particularly under low-intensity attacks, where accuracy dropped to 0.64 at $\epsilon = 0.01$ for FGSM, and to 0.76 at $\theta = 0.1$, $\gamma = 0.4$ for JSMA. The effectiveness of three defence mechanisms was subsequently evaluated: Adversarial Training, a Hybrid Autoencoder-Random Forest Classifier, and a Feature Partition Forest with a Random Forest base (FPF-RF). While adversarial training effectively restored performance, the FPF-RF remained the most resilient. It demonstrated superior and consistent robustness across all attack parameters and even outperforming the baseline under intense JSMA perturbations (achieving an F1-score of 0.90). This work concludes that a combination of strategic model selection and specifically tailored defence architectures is essential for enhancing adversarial robustness in ICS environments. It also provides a critical step towards securing critical infrastructure against evolving cyber threats.