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

This paper tests whether data augmentation, specifically the addition of image perturbations to training data, can improve the robustness of image-classification models to common perturbations they encounter at test time. Using the CIFAR-10 dataset, the study applies seven distinct perturbations to training images and then measures how accurately the resulting model classifies unseen test data, concluding that perturbation-based augmentation does improve robust accuracy. The work frames robustness as more than a technical metric. It opens with the problem of bias in AI systems, citing research that found bias in a substantial share of the 'facts' used by AI models, and argues that brittle classifiers which fail on underrepresented or out-of-distribution inputs can produce discriminatory real-world outcomes such as racial profiling or sexism. Improving a model's ability to generalise to perturbed and unfamiliar images is therefore presented as both an engineering goal and a fairness goal. The paper explains the supervised-learning basis of image classifiers, the different sources of bias (data, algorithmic and human, along with selection, framing and label bias), and the concept of adversarial examples, perturbations engineered to be imperceptible yet capable of fooling a model. Its focus is on demonstrating, through controlled experiments with convolutional neural networks, that deliberately exposing a model to perturbed images during training makes it more reliable afterward. It connects this to practical applications including safer autonomous vehicles and more capable surveillance and object-tracking systems, positioning data augmentation as a accessible route to fairer and more dependable computer-vision models.