Detection of sticker based adversarial attacks

摘要

Adversarial examples revealed and important aspect of convolutional neural networks and are getting more and more attention in machine learning. It was shown that not only small perturbations, covering the whole image can be applied but also sticker based attacks, concentrated on small regions of the image can cause misclassification. Meanwhile the first type of attack is theoretical the later can be applied in practice and lead tomisclassification in image processing pipelines. In this paper we show a method how sticker based adversarial samples can be detected by calculating the responses of the neurons in the last layers and estimating the measure of region based classification consistency.