A Convolutional Neural Network (CNN) is one branch of Deep Learning widely used for image classification. CNN have complex architectures and capable of achieving high accuracy and producing good results. However, CNN have limitations, especially when dealing with noisy images. Image noise can decrease classification performance and increase network training time. This research was tested the robustness of two CNN methods, namely VGG16 and ResNet50, in processing images with added Gaussian noise at various levels, without any preprocessing. The dataset used is the Rice Image Dataset, which consists of images of 5 different types of rice. The data is divided into three parts: 70% for training, 20% for testing, and 10% for validation. The results of this study show that as the variability in generating Gaussian noise in the images increases, the loss function value consistently increases, and the accuracy decreases. However, the increase in the loss function value and the decrease in accuracy are not significantly different among the different levels of noise variability.