Quantum machine learning (QML) has emerged as a promising paradigm for intrusion detection systems (IDS) in resource-constrained and highly dynamic IoT networks due to its superior data efficiency and ability to capture complex non-linear attack patterns. However, the adversarial robustness of QML-based IDS remains largely unexplored, posing significant security risks in real-world deployments where attackers can craft subtle perturbations against flow-ratio features. This paper presents the first comprehensive adversarial robustness benchmark of two representative Quantum Learning-based IDS models, namely Quantum Support Vector Classifier (QSVC) and Quantum Convolutional Neural Network (QCNN), against their classical counterparts Support Vector Classifier (SVC) and Convolutional Neural Network (CNN). Using the ACI IoT dataset, we evaluate the models under distinct threat configurations: white-box perturbations via the Fast Gradient Sign Method (FGSM), and comprehensive black-box evaluations using Zeroth Order Optimization (ZOO) and HopSkipJump (HSJ). Extensive experiments reveal a highly attack-dependent robustness profile: quantum models demonstrate distinct resilience against boundary-search black-box attacks (HSJ), yet exhibit heightened vulnerability to score-based probing (ZOO). Furthermore, under gradient-based white-box attacks (FGSM), robustness is architecture-specific, benefiting quantum kernel methods (QSVC) while challenging variational neural approaches (QCNN). These findings highlight both the potential advantages and vulnerabilities of quantum kernel and quantum neural approaches in IoT security, providing essential guidelines for designing trustworthy QML-based IDS in 5G/6G and Noisy Intermediate-Scale Quantum (NISQ)-era networks.