Deep learning models, particularly transformer-based architectures, are highly susceptible to adversarial attacks. These attacks introduce small, deliberate perturbations to inputs that do not alter their meaning but significantly degrade model performance. This paper examines the impact of adversarial examples on the robustness of two transformer-based models. The study begins with pretraining the ELECTRA-small and DeBERTa models on a standard question-and-answer (QA) dataset, followed by evaluation on multiple adversarial datasets. This process helps identify key sources of performance degradation in the models under various adversarial attack patterns. To enhance robustness, an inoculation process is applied, where the pretrained model (referred to as the patient) is further trained on a small subset (referred to as the pathogen) of the adversarial datasets. Findings are reported and compared between the inoculated models and the pretrained ones. Results show a general narrowing of the performance gap in the retrained models on similar adversarial datasets. However, a notable decline in performance is observed on the original SQuAD dataset used for pretraining. These findings highlight the trade-offs involved in improving model resilience to adversarial attacks while maintaining performance on standard datasets.