Despite significant advancements in deep learning, adversarial attacks remain a critical challenge, threatening the robustness and security of neural networks. Current models, including convolutional neural networks (CNNs), are still vulnerable to these attacks, and the development of effective defenses remains an open problem. While previous studies have primarily focused on transferability between models with different backbones, in this paper, we investigate the transferability and effectiveness of adversarial attacks between standard CNN architectures and dilated CNN architectures, as dilation is known for enhancing the efficiency of CNNs. Specifically, we evaluate two critical aspects: (1) whether dilated CNNs, when used defensively, exhibit enhanced resistance to adversarial attacks and (2) whether adversarial samples generated using dilated CNNs result in weaker or less transferable attacks. Using the Fast Gradient Sign Method (FGSM), we generate adversarial samples for multiple CNN and dilated CNN models, analyzing both direct attacks (targeting the same model) and cross attacks (targeting different architectures). Additionally, we vary the FGSM epsilon parameter to assess the impact of different attack strengths. The results of the evaluation indicate that dilated CNNs demonstrate partial resistance to attacks from standard CNNs, especially at higher dilation rates, but remain vulnerable overall. Conversely, adversarial samples generated using dilated CNNs are less effective when transferred to standard CNNs, highlighting a trade-off between efficiency, robustness, and adversarial susceptibility.