Explaining Image Misclassification in Deep Learning via Adversarial Examples

Identificador:  imarina:9229338

Handle: https://hdl.handle.net/20.500.11797/imarina9229338

Autores:  Haffar, Rami; Jebreel, Najeeb Moharram; Domingo-Ferrer, Josep; Sanchez, David

Resumen:
With the increasing use of convolutional neural networks (CNNs) for computer vision and other artificial intelligence tasks, the need arises to interpret their predictions. In this work, we tackle the problem of explaining CNN misclassification of images. We propose to construct adversarial examples that allow identifying the regions of the input images that had the largest impact on the CNN wrong predictions. More specifically, for each image that was incorrectly classified by the CNN, we implemented an inverted adversarial attack consisting on modifying the input image as little as possible so that it becomes correctly classified. The changes made to the image to fix classification errors explain the causes of misclassification and allow adjusting the model and the data set to obtain more accurate models. We present two methods, of which the first one employs the gradients from the CNN itself to create the adversarial examples and is meant for model developers. However, end users only have access to the CNN model as a black box. Our second method is intended for end users and employs a surrogate model to estimate the gradients of the original CNN model, which are then used to create the adversarial examples. In our experiments, the first method achieved 99.67% success rate at finding the misclassification explanations and needed on average 1.96 queries per misclassified image to build the corresponding adversarial example. The second method achieved 73.08% success rate at finding the explanations with 8.73 queries per image on average.