Comparative Study of Adversarial Image Attacks

Adversarial attacks exploit subtle input modifications to mislead deep neural networks (DNNs), posing significant risks to critical applications like autonomous vehicles and medical imaging. This study examines adversarial image generation using methods such as Fast Gradient Sign Method (FGSM), DeepFool, and Carlini & Wagner (C&W) attacks to assess DNN robustness. This approach involves generating adversarial samples with controlled perturbations and evaluating their impact on model performance. The findings contribute to a deeper understanding of adversarial risks and pave the way for more resilient AI systems capable of withstanding deceptive inputs in real-world applications. The results highlight FGSM's efficiency in rapid perturbation generation, DeepFool's minimal iterative modifications, and C& W's optimized, imperceptible attacks, all of which significantly reduce model accuracy