Using Data Augmentation to Improve the Robustness of Image Classification Models to Common Perturbations

This study identified the improvement of robust accuracy of image classification models by using image perturbations. Seven perturbations were added to images in the CIFAR-10 dataset, after which the model was tested for its accuracy in identifying images in test data. The research confirms that using perturbations can improve the robustness of image classification models. This paper emphasizes understanding and evaluating how image perturbations added to the training dataset allow Image Classifiers to perform better on test data. This method can lead to fewer discriminatory outcomes when employed in real-world applications, reducing concerns such as racism and sexism. Applications include more efficient self-driving cars and improved object tracking in CCTV systems.