This paper presents a one-pixel adversarial attack method for fooling deep neural networks (DNNs) using differential evolution (DE). The method modifies only one pixel in an input image to alter the DNN's classification output. It requires minimal information (black-box attack) and can fool various types of networks due to DE's inherent properties. Experiments show that 67.97% of Kaggle CIFAR-10 test images and 16.04% of ImageNet test images can be perturbed to target classes with high confidence. The attack is effective in extreme limited scenarios, revealing vulnerabilities in DNNs to low-dimensional attacks. The method also demonstrates the application of DE in adversarial machine learning for generating low-cost adversarial attacks to evaluate network robustness. The attack is evaluated on CIFAR-10 and ImageNet datasets, showing high success rates for targeted and non-targeted attacks. Results indicate that one-pixel attacks can fool multiple classes and are effective across different network structures. The method is compared with random attacks, showing superior performance in terms of attack accuracy and confidence. The study highlights the vulnerability of DNNs to adversarial perturbations and the potential of evolutionary computation in adversarial machine learning. Future work includes exploring more advanced algorithms and extending the attack to other domains like natural language processing and speech recognition.This paper presents a one-pixel adversarial attack method for fooling deep neural networks (DNNs) using differential evolution (DE). The method modifies only one pixel in an input image to alter the DNN's classification output. It requires minimal information (black-box attack) and can fool various types of networks due to DE's inherent properties. Experiments show that 67.97% of Kaggle CIFAR-10 test images and 16.04% of ImageNet test images can be perturbed to target classes with high confidence. The attack is effective in extreme limited scenarios, revealing vulnerabilities in DNNs to low-dimensional attacks. The method also demonstrates the application of DE in adversarial machine learning for generating low-cost adversarial attacks to evaluate network robustness. The attack is evaluated on CIFAR-10 and ImageNet datasets, showing high success rates for targeted and non-targeted attacks. Results indicate that one-pixel attacks can fool multiple classes and are effective across different network structures. The method is compared with random attacks, showing superior performance in terms of attack accuracy and confidence. The study highlights the vulnerability of DNNs to adversarial perturbations and the potential of evolutionary computation in adversarial machine learning. Future work includes exploring more advanced algorithms and extending the attack to other domains like natural language processing and speech recognition.