Deep learning has made a major progress and changes to medical imaging in particularly identifying diseases such as pneumonia from chest X-rays. These models are vulnerable to adversarial attacks, small or almost invisible changes in input images that can mislead the systems and it is hard to predict. These weaknesses show a real concern in clinical use where diagnostic accuracy is important. In this study, we propose a defense method that increases the robustness of deep learning models against these adversarial attacks without reducing their diagnostic capability. Our method employs EfficientNetB0, a compact but effective neural network which is trained on chest X-ray data to differentiate whether the image is pneumonia or normal image. To see a model’s flexibility, we generate adversarial samples using the Fast Gradient Sign Method (FGSM) and after that, we apply a defense mechanism to minimize their impact through Total Variation Minimization (TVM), which removes noise and retains original image features. The final results show that combining EfficientNetB0 with TVM strengthens accuracy under attack and maintains strong performance on clean data. This work shows a practical path toward secure and reliable deep learning-based medical image analysis and provides information about how much adversarial defense is important in healthcare systems.