Deep Learning (DL) models have revolutionized critical fields like healthcare and autonomous systems but remain vulnerable to adversarial attacks, where minor perturbations in input data cause incorrect predictions. This project addresses these challenges by implementing adversarial training, a method that strengthens model robustness by incorporating adversarial examples crafted using Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), and Carlini and Wagner (CW) attack into the training process. The results show that adversarially trained models achieve significant resilience, maintaining high accuracy on both clean and adversarial datasets. The ResNet50 model achieved 93.12% accuracy on clean images before adversarial training, and after adversarial training, the accuracy increased to 95.0%, and the model’s performance improved from 43.29 to 93.6% under adversarial attacks. This approach ensures the reliability and robustness of DL models in safety-critical applications while maintaining computational efficiency.