Sensitivity Of Convolutional Neural Networks Research Articles

Cognitive data augmentation for adversarial defense via pixel masking

• We proposed PixelMask: a novel data augmentation technique for adversarial robustness. • Extensive experiments showcase the effectiveness of the proposed defense algorithm. • Proposed PixelMask defense outperforms other strong data augmentation techniques. The vulnerability of deep networks towards adversarial perturbations has motivated the researchers to design detection and mitigation algorithms. Inspired by the dropout and dropconnect algorithms as well as augmentation techniques, this paper presents “PixelMask” based data augmentation as an efficient method of reducing the sensitivity of convolutional neural networks (CNNs) towards adversarial attacks. In the proposed approach, samples generated using PixelMask are used as augmented data, which helps in learning robust CNN models. Experiments performed with multiple databases and architectures show that the proposed PixelMask based data augmentation approach improves the classification performance on adversarially perturbed images. The proposed defense mechanism can be applied effectively for different adversarial attacks and can easily be combined with any deep neural network (DNN) architecture to increase the robustness. The effectiveness of the proposed defense is demonstrated in gray-box, white-box, and unseen train-test attack scenarios. For example, on the CIFAR-10 database under adaptive attack (i.e., projected gradient descent), the proposed PixelMask is able to improve the recognition performance of CNN by at-least 22.69%. Another advantage of the proposed algorithm over several existing defense algorithms is that the proposed defense either is able to retain or increase the classification accuracy of clean examples.

Absum: Simple Regularization Method for Reducing Structural Sensitivity of Convolutional Neural Networks

We propose Absum, which is a regularization method for improving adversarial robustness of convolutional neural networks (CNNs). Although CNNs can accurately recognize images, recent studies have shown that the convolution operations in CNNs commonly have structural sensitivity to specific noise composed of Fourier basis functions. By exploiting this sensitivity, they proposed a simple black-box adversarial attack: Single Fourier attack. To reduce structural sensitivity, we can use regularization of convolution filter weights since the sensitivity of linear transform can be assessed by the norm of the weights. However, standard regularization methods can prevent minimization of the loss function because they impose a tight constraint for obtaining high robustness. To solve this problem, Absum imposes a loose constraint; it penalizes the absolute values of the summation of the parameters in the convolution layers. Absum can improve robustness against single Fourier attack while being as simple and efficient as standard regularization methods (e.g., weight decay and L1 regularization). Our experiments demonstrate that Absum improves robustness against single Fourier attack more than standard regularization methods. Furthermore, we reveal that robust CNNs with Absum are more robust against transferred attacks due to decreasing the common sensitivity and against high-frequency noise than standard regularization methods. We also reveal that Absum can improve robustness against gradient-based attacks (projected gradient descent) when used with adversarial training.

Characterizing robustness and sensitivity of convolutional neural networks for quantitative analysis of mitochondrial morphology

BackgroundQuantitative analysis of mitochondrial morphology plays important roles in studies of mitochondrial biology. The analysis depends critically on segmentation of mitochondria, the image analysis process of extracting mitochondrial morphology from images. The main goal of this study is to characterize the performance of convolutional neural networks (CNNs) in segmentation of mitochondria from fluorescence microscopy images. Recently, CNNs have achieved remarkable success in challenging image segmentation tasks in several disciplines. So far, however, our knowledge of their performance in segmenting biological images remains limited. In particular, we know little about their robustness, which defines their capability of segmenting biological images of different conditions, and their sensitivity, which defines their capability of detecting subtle morphological changes of biological objects.MethodsWe have developed a method that uses realistic synthetic images of different conditions to characterize the robustness and sensitivity of CNNs in segmentation of mitochondria. Using this method, we compared performance of two widely adopted CNNs: the fully convolutional network (FCN) and the U‐Net. We further compared the two networks against the adaptive active‐mask (AAM) algorithm, a representative of high‐performance conventional segmentation algorithms.ResultsThe FCN and the U‐Net consistently outperformed the AAM in accuracy, robustness, and sensitivity, often by a significant margin. The U‐Net provided overall the best performance.ConclusionsOur study demonstrates superior performance of the U‐Net and the FCN in segmentation of mitochondria. It also provides quantitative measurements of the robustness and sensitivity of these networks that are essential to their applications in quantitative analysis of mitochondrial morphology.