A Suite of Incremental Image Degradation Operators for Testing Image Classification Algorithms

Convolutional Neural Networks (CNN) are extremely popular for modelling sound and images, but they suffer from a lack of robustness that could threaten their usefulness in applications where reliability is important. Recent studies have shown how it is possible to maliciously create adversarial images—those that appear to the human observer as perfect examples of one class but that fool a CNN into assigning them to a different, incorrect class. It takes some effort to make these images as they need to be designed specifically to fool a given network. In this paper we show that images can be degraded in a number of simple ways that do not need careful design and that would not affect the ability of a human observer, but which cause severe deterioration in the performance of three different CNN models. We call the speed of the deterioration in performance due to incremental degradations in image quality the degradation profile of a model and argue that reporting the degradation profile is as important as reporting performance on clean images