Technological advancements are reshaping our world, bringing with them new cyber threats, including sophisticated adversarial attacks on machine learning models. This research focuses on two types of these attacks: clean-label poisoning and decision-based attacks, which threaten the integrity of models. By utilizing two prominent datasets, MNIST and CIFAR-10, the study performs a comparative analysis to assess how these attacks impact neural network models tailored for image classification. Employing black-box attack techniques, the research simulates realistic conditions that attackers might use. The implementation and evaluation of these models are conducted using TensorFlow, focusing on the effects of attacks on model accuracy and exploring potential defense mechanisms. The objective is to understand the severity of these adversarial attacks and to identify effective defenses to protect machine learning models from such vulnerabilities.