CNN Vulnerability On Untargeted Adversarial Attacks

Convolutional Neural Networks (CNNs) are widely used in various AI-driven industries, including healthcare, autonomous systems, and cybersecurity, due to their high accuracy in image classification. However, these models are vulnerable to adversarial attacks, where small, imperceptible perturbations in input data can cause incorrect predictions. This study evaluates the impact of five adversarial attacks: Fast Gradient Sign Method (FGSM), Iterative FGSM (IFGSM), Momentum Iterative FGSM (MIFGSM), Deepfool, and OnePixel Attack on the accuracy of a CNN trained on the CIFAR-10 dataset. Each attack is analyzed for its unique influence on model performance, revealing significant accuracy drops under adversarial conditions. The results are analyzed and illustrate the potential security risks of adversarial attacks, underscoring the need for robust defenses in CNN based systems.