VATRA helps machine learning practitioners compare how models perform on ordinary images and images altered by adversarial attacks. Training a model to resist attacks can affect its accuracy and how evenly it performs across different classes. We developed VATRA through an iterative design process to make these trade-offs easier to explore, from overall model performance down to individual examples.