The computer-aided system to be used to detect Alzheimer disease (AD) should have computationally efficient models that are resistant to distortions that usually occur in medical imaging. This work presents a lightweight model of deep learning that is robustness-based and has the ability to classify AD by stage using structural MRI data. The given framework is based on the EfficientNetB0 backbone because it has a compound scaling property to enable efficient features learning with lower training complexity. The label-conscious augmentation and adversarial learning through the Fast Gradient Sign Method (FGSM) is also optimized to give a better resistance to imperceptible perturbation. Experimental assessment illustrates a test accuracy of 98.36% percent and an adversarial accuracy of 99.43% under Fast Gradient Sign Method (FGSM)-based adversarial evaluation through four stages of AD, namely, Non-Demented, Very Mild-Demented, Mild-Demented and Moderate-Demented. These findings suggest that the proposed solution is capable of high generalization, minimal computation cost, and high resistance to adversarial changes and would be an effective solution to clinical decision support in a resource-limited healthcare environment.