Robustness of Convolutional Neural Networks to Partial Test-Set Image Corruption on Fashion-MNIST

Image classifiers often report high accuracy on clean benchmark data, but in the real world, their inputs are not always clean. This paper tests how two convolutional models respond when part of the fashion-MNIST test set is deliberately corrupted. A custom four-layer convolutional neural network and an ImageNet-pretrained ResNet-18 were trained only on clean Fashion-MNIST images. A fixed corruption, Gaussian blur followed by contrast enhancement, was then applied to 0%, 10%, 30%, 50%, 70%, and 100% of the official test images. Both models used the same corrupted image set. Accuracy, mean predictive entropy, and t-distributed stochastic neighbor embedding plots were used to compare the runs. The custom model reached 91.31% accuracy on clean data but fell to 10.82% when every test image was corrupted. ResNet-18 started lower, at 88.10%, but reached 20.44% at 100% corruption. Entropy also rose with the corrupted percentage, from 0.1729 to 0.5379 for the custom model and from 0.1517 to 0.3073 for ResNet-18. These suggest that both clean-trained models were fragile under this specific corruption, while ResNet-18 held up better in this run. A possible explanation is that its deeper residual structure and ImageNet-pretrained layers preserved more coarse shape information after blur weakened fine local details.