Fast Gradient Sign Method (FGSM) Variants in White Box Settings: A Comparative Study

Adversarial Machine Learning is a critical research area focused on understanding the vulnerabilities of machine learning models. Adversarial attacks have the potential to significantly impact deep learning models, particularly Convolutional Neural Networks (CNNs), undermining their reliability and real-world applications. This research study analyzes the Fast Gradient Sign Method (FGSM) and its variants, examining their effects on the ResNet-50 CNN model. By leveraging the model's training on the ImageNet dataset, this study aims to reveal the mechanisms of each attack and identify their weaknesses. Through a comprehensive comparison and analysis of their behaviors, this research study aims to provide developers with a thorough understanding of FGSM's characteristics, thereby facilitating the development of robust defense strategies against adversarial threats in deep learning models.