The range of applications of Neural Networks encompasses image classification. However, Neural Networks are vulnerable to attacks, and may misclassify adversarial images, leading to potentially disastrous consequences. Pursuing some of our previous work, we provide an extended proof of concept of a black-box, targeted, non-parametric attack using evolutionary algorithms to fool both Neural Networks and humans at the task of image classification. Our feasibility study is performed on VGG-16 trained on CIFAR-10. For any category of CIFAR-10, one chooses an image classified by VGG-16 as belonging to . From there, two scenarios are addressed. In the first scenario, a target category is fixed a priori. We construct an evolutionary algorithm that evolves to a modified image that VGG-16 classifies as belonging to . In the second scenario, we construct another evolutionary algorithm that evolves to a modified image that VGG-16 is unable to classify. In both scenarios, the obtained adversarial images remain so close to the original one that a human would likely classify them as still belonging to .
Recently, convolutional neural networks (CNNs) have become the main drivers in many image recognition applications. However, they are vulnerable to adversarial attacks, which can lead to disastrous consequences. This paper introduces ShuffleDetect as a new and efficient unsupervised method for the detection of adversarial images against trained convolutional neural networks. Its main feature is to split an input image into non-overlapping patches, then swap the patches according to permutations, and count the number of permutations for which the CNN classifies the unshuffled input image and the shuffled image into different categories. The image is declared adversarial if and only if the proportion of such permutations exceeds a certain threshold value. A series of 8 targeted or untargeted attacks was applied on 10 diverse and state-of-the-art ImageNet-trained CNNs, leading to 9500 relevant clean and adversarial images. We assessed the performance of ShuffleDetect intrinsically and compared it with another detector. Experiments show that ShuffleDetect is an easy-to-implement, very fast, and near memory-free detector that achieves high detection rates and low false positive rates.
31 October 2022 in Journal of Information and Telecommunication
To perform image recognition, Convolutional Neural Networks (CNNs) assess any image by first resizing it to its input size. In particular, high-resolution images are scaled down, say to for CNNs trained on ImageNet. So far, existing attacks, aiming at creating an adversarial image that a CNN would misclassify while a human would not notice any difference between the modified and unmodified images, proceed by creating adversarial noise in the resized domain and not in the high-resolution domain. The complexity of directly attacking high-resolution images leads to challenges in terms of speed, adversity and visual quality, making these attacks infeasible in practice. We design an indirect attack strategy that lifts to the high-resolution domain any existing attack that works efficiently in the CNN's input size domain. Adversarial noise created via this method is of the same size as the original image. We apply this approach to 10 state-of-the-art CNNs trained on ImageNet, with an evolutionary algorithm-based attack. Our method succeeded in 900 out of 1000 trials to create such adversarial images, that CNNs classify with probability in the adversarial category. Our indirect attack is the first effective method at creating adversarial images in the high-resolution domain.
Through the addition of humanly imperceptible noise to an image classified as belonging to a category , targeted adversarial attacks can lead convolutional neural networks (CNNs) to classify a modified image as belonging to any predefined target class . To achieve a better understanding of the inner workings of adversarial attacks, this study analyzes the adversarial images created by two completely opposite attacks against 10 ImageNet-trained CNNs. A total of adversarial images are created by , a black-box evolutionary algorithm (EA), and by the basic iterative method (BIM), a white-box, gradient-based attack. We inspect and compare these two sets of adversarial images from different perspectives: the behavior of CNNs at smaller image regions, the image noise frequency, the adversarial image transferability, the image texture change, and penultimate CNN layer activations. We find that texture change is a side effect rather than a means for the attacks and that -relevant features only build up significantly from image regions of size onwards. In the penultimate CNN layers, both attacks increase the activation of units that are positively related to and units that are negatively related to . In contrast to ’s white noise nature, BIM predominantly introduces low-frequency noise. BIM affects the original features more than , thus producing slightly more transferable adversarial images. However, the transferability with both attacks is low, since the attacks’ -related information is specific to the output layers of the targeted CNN. We find that the adversarial images are actually more transferable at regions with sizes of than at full scale.