Blerta Lindqvist

Blerta Lindqvist

This paper uses symmetry to make Convolutional Neural Network classifiers (CNNs) robust against adversarial perturbation attacks. Such attacks add perturbation to original images to generate adversarial images that fool classifiers such as road sign classifiers of autonomous vehicles. Although symmetry is a pervasive aspect of the natural world, CNNs are unable to handle symmetry well. For example, a CNN can classify an image differently from its mirror image. For an adversarial image that misclassifies with a wrong label $l_w$, CNN inability to handle symmetry means that a symmetric adversarial image can classify differently from the wrong label $l_w$. Further than that, we find that the classification of a symmetric adversarial image reverts to the correct label. To classify an image when adversaries are unaware of the defense, we apply symmetry to the image and use the classification label of the symmetric image. To classify an image when adversaries are aware of the defense, we use mirror symmetry and pixel inversion symmetry to form a symmetry group. We apply all the group symmetries to the image and decide on the output label based on the agreement of any two of the classification labels of the symmetry images. Adaptive attacks fail because they need to rely on loss functions that use conflicting CNN output values for symmetric images. Without attack knowledge, the proposed symmetry defense succeeds against both gradient-based and random-search attacks, with up to near-default accuracies for ImageNet. The defense even improves the classification accuracy of original images.

Blerta Lindqvist

We examine whether symmetry can be used to defend tree-based ensemble classifiers such as gradient-boosting decision trees (GBDTs) against adversarial perturbation attacks. The idea is based on a recent symmetry defense for convolutional neural network classifiers (CNNs) that utilizes CNNs' lack of invariance with respect to symmetries. CNNs lack invariance because they can classify a symmetric sample, such as a horizontally flipped image, differently from the original sample. CNNs' lack of invariance also means that CNNs can classify symmetric adversarial samples differently from the incorrect classification of adversarial samples. Using CNNs' lack of invariance, the recent CNN symmetry defense has shown that the classification of symmetric adversarial samples reverts to the correct sample classification. In order to apply the same symmetry defense to GBDTs, we examine GBDT invariance and are the first to show that GBDTs also lack invariance with respect to symmetries. We apply and evaluate the GBDT symmetry defense for nine datasets against six perturbation attacks with a threat model that ranges from zero-knowledge to perfect-knowledge adversaries. Using the feature inversion symmetry against zero-knowledge adversaries, we achieve up to 100% accuracy on adversarial samples even when default and robust classifiers have 0% accuracy. Using the feature inversion and horizontal flip symmetries against perfect-knowledge adversaries, we achieve up to over 95% accuracy on adversarial samples for the GBDT classifier of the F-MNIST dataset even when default and robust classifiers have 0% accuracy.

Blerta Lindqvist

Despite extensive research into adversarial attacks, we do not know how adversarial attacks affect image pixels. Knowing how image pixels are affected by adversarial attacks has the potential to lead us to better adversarial defenses. Motivated by instances that we find where strong attacks do not transfer, we delve into adversarial examples at pixel level to scrutinize how adversarial attacks affect image pixel values. We consider several ImageNet architectures, InceptionV3, VGG19 and ResNet50, as well as several strong attacks. We find that attacks can have different effects at pixel level depending on classifier architecture. In particular, input pre-processing plays a previously overlooked role in the effect that attacks have on pixels. Based on the insights of pixel-level examination, we find new ways to detect some of the strongest current attacks.

Blerta Lindqvist

Neural network classifiers are vulnerable to misclassification of adversarial samples, for which the current best defense trains classifiers with adversarial samples. However, adversarial samples are not optimal for steering attack convergence, based on the minimization at the core of adversarial attacks. The minimization perturbation term can be minimized towards $0$ by replacing adversarial samples in training with duplicated original samples, labeled differently only for training. Using only original samples, Target Training eliminates the need to generate adversarial samples for training against all attacks that minimize perturbation. In low-capacity classifiers and without using adversarial samples, Target Training exceeds both default CIFAR10 accuracy ($84.3$%) and current best defense accuracy (below $25$%) with $84.8$% against CW-L$_2$($κ=0$) attack, and $86.6$% against DeepFool. Using adversarial samples against attacks that do not minimize perturbation, Target Training exceeds current best defense ($69.1$%) with $76.4$% against CW-L$_2$($κ=40$) in CIFAR10.

Blerta Lindqvist

Recent adversarial defense approaches have failed. Untargeted gradient-based attacks cause classifiers to choose any wrong class. Our novel white-box defense tricks untargeted attacks into becoming attacks targeted at designated target classes. From these target classes, we can derive the real classes. Our Target Training defense tricks the minimization at the core of untargeted, gradient-based adversarial attacks: minimize the sum of (1) perturbation and (2) classifier adversarial loss. Target Training changes the classifier minimally, and trains it with additional duplicated points (at 0 distance) labeled with designated classes. These differently-labeled duplicated samples minimize both terms (1) and (2) of the minimization, steering attack convergence to samples of designated classes, from which correct classification is derived. Importantly, Target Training eliminates the need to know the attack and the overhead of generating adversarial samples of attacks that minimize perturbations. We obtain an 86.2% accuracy for CW-L2 (confidence=0) in CIFAR10, exceeding even unsecured classifier accuracy on non-adversarial samples. Target Training presents a fundamental change in adversarial defense strategy.

Blerta Lindqvist, Rauf Izmailov

State-of-the-art adversarial attacks are aimed at neural network classifiers. By default, neural networks use gradient descent to minimize their loss function. The gradient of a classifier's loss function is used by gradient-based adversarial attacks to generate adversarially perturbed images. We pose the question whether another type of optimization could give neural network classifiers an edge. Here, we introduce a novel approach that uses minimax optimization to foil gradient-based adversarial attacks. Our minimax classifier is the discriminator of a generative adversarial network (GAN) that plays a minimax game with the GAN generator. In addition, our GAN generator projects all points onto a manifold that is different from the original manifold since the original manifold might be the cause of adversarial attacks. To measure the performance of our minimax defense, we use adversarial attacks - Carlini Wagner (CW), DeepFool, Fast Gradient Sign Method (FGSM) - on three datasets: MNIST, CIFAR-10 and German Traffic Sign (TRAFFIC). Against CW attacks, our minimax defense achieves 98.07% (MNIST-default 98.93%), 73.90% (CIFAR-10-default 83.14%) and 94.54% (TRAFFIC-default 96.97%). Against DeepFool attacks, our minimax defense achieves 98.87% (MNIST), 76.61% (CIFAR-10) and 94.57% (TRAFFIC). Against FGSM attacks, we achieve 97.01% (MNIST), 76.79% (CIFAR-10) and 81.41% (TRAFFIC). Our Minimax adversarial approach presents a significant shift in defense strategy for neural network classifiers.

Blerta Lindqvist, Shridatt Sugrim, Rauf Izmailov

Classifiers fail to classify correctly input images that have been purposefully and imperceptibly perturbed to cause misclassification. This susceptability has been shown to be consistent across classifiers, regardless of their type, architecture or parameters. Common defenses against adversarial attacks modify the classifer boundary by training on additional adversarial examples created in various ways. In this paper, we introduce AutoGAN, which counters adversarial attacks by enhancing the lower-dimensional manifold defined by the training data and by projecting perturbed data points onto it. AutoGAN mitigates the need for knowing the attack type and magnitude as well as the need for having adversarial samples of the attack. Our approach uses a Generative Adversarial Network (GAN) with an autoencoder generator and a discriminator that also serves as a classifier. We test AutoGAN against adversarial samples generated with state-of-the-art Fast Gradient Sign Method (FGSM) as well as samples generated with random Gaussian noise, both using the MNIST dataset. For different magnitudes of perturbation in training and testing, AutoGAN can surpass the accuracy of FGSM method by up to 25\% points on samples perturbed using FGSM. Without an augmented training dataset, AutoGAN achieves an accuracy of 89\% compared to 1\% achieved by FGSM method on FGSM testing adversarial samples.