Quantum classifiers have recently been found to be vulnerable to adversarial attacks. These attacks perturb the input data to mislead the classifier to make incorrect predictions. Adversarial training has since emerged as a promising defence strategy for quantum classifiers. In adversarial training, the training set consists of both clean and adversarially perturbed data, where the latter are generated assuming a known adversarial attack. Thus, the classifier is trained to minimize an adversarial loss, which quantifies the worst loss suffered by the classifier due to adversarial perturbations. Although quantum classifiers trained in this way incur lower adversarial loss on the training dataset, there is still limited understanding of how well these classifiers generalize to new, previously unseen, adversarially perturbed data. This chapter reviews recent research focused on understanding the generalization performance of adversarially trained quantum classifiers. To this end, we first introduce the learning-theoretic concept of Rademacher complexity and demonstrate how it can quantify the generalization performance of standard, non-adversarially trained, quantum classifiers. We then extend this framework to the adversarial setting considering different types of adversarial attacks, including classical adversarial attacks perturbing the classical input data before it is embedded into a quantum state, and quantum adversarial attacks perturbing the quantum states directly. For classical attacks, we also discuss the impact of the quantum embedding on the generalization performance of adversarially trained quantum classifiers.