A Defensive Approach against Adversarial Examples Based on Manifold Learning

Deep neural network has been shown to be vulnerable to adversarial examples. It means that a well-trained network can be fooled to make wrong predictions by adding imperceptible changes to clean images. Research has found that adversarial examples mostly change the irrelevant features which have less impact on classification. In this paper, a new approach is proposed to defend adversarial examples by masking the irrelevant features based on manifold learning. Specially, Isomap is used for dimensionality reduction, which is one of the manifold learning algorithms. Then the approach analyzes the respective correlations between the result and each feature, and filters the features with lower correlation coefficient. Compared with other defensive approaches, the proposed method is easy to implement and does not rely on adversarial examples. Through experiments on MNIST, CIFAR-10, CIFAR-100 and our own traffic sign dataset, it shows that the proposed method can increase the robustness of classifiers against adversarial examples effectively.