In recent years, the rapid advancement of artificial intelligence and the growing volume of data have made distributed systems essential for addressing complex technical challenges and processing large-scale datasets. Neural Architecture Search (NAS) has emerged as a promising solution for the automatic discovery of high-performance deep neural networks (DNNs), particularly in the design of sophisticated convolutional neural networks (CNNs). This paper presents a NAS method based on a heterogeneous CPU + GPU distributed system, utilizing genetic algorithms (GA) for the automatic search of optimal CNN architectures. By encoding the core concepts of ResNet, WideResNet, DenseNet, and MobileNet into GA individuals, our approach enhances the diversity and performance of CNN architectures. Experimental results demonstrate that our model achieves nearly a 13% improvement in accuracy on the CIFAR-10 dataset, outperforming traditional models such as ResNet and DenseNet. Additionally, it significantly reduces model parameters by up to 80%, making it more suitable for edge computing environments.