This research evaluates the robustness of DenseNet-161 neural network architecture using the Stanford Dogs dataset for fine-grained image classification. Specifically, it examines the model’s performance against adversarial examples developed with the Fast Gradient Sign Method and Projected Gradient Descent attack techniques . These attacks created inputs that appear normal to humans but induce errors in machine learning models. In this process, adversarial perturbations were regulated with \(L_1\) , \(L_2\) , and \(L_{\infty }\) norms to ensure subtle modification. Although the DenseNet-161 model successfully provided a high accuracy of 83.75% on clean (without attack) Stanford Dogs data, it was severely impacted by adversarial attacks, which detrimentally affected confidence scores and predictions. Hence, the model was highly sensitive to uniform pixel-wise perturbations from \(L_{\infty }\) norm attacks. The results reveal vulnerabilities in current artificial intelligence models and emphasize the importance of adversarial robustness, especially in security-sensitive applications. This analysis provides insights into developing broader defenses across perturbation norms rather than solely relying on \(L_{\infty }\) robustness, bridging gaps in understanding model limitations and advancing toward more reliable systems.