English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Deep neural networks (DNNs) are known to be vulnerable to adversarialexamples that are crafted with imperceptible perturbations, i.e., a smallchange in an input image can induce a mis-classification, and thus threatensthe reliability of deep learning based deployment systems. Adversarial training(AT) is often adopted to improve robustness through training a mixture ofcorrupted and clean data. However, most of AT based methods are ineffective indealing with transferred adversarial examples which are generated to fool awide spectrum of defense models, and thus cannot satisfy the generalizationrequirement raised in real-world scenarios. Moreover, adversarially training adefense model in general cannot produce interpretable predictions towards theinputs with perturbations, whilst a highly interpretable robust model isrequired by different domain experts to understand the behaviour of a DNN. Inthis work, we propose a novel approach based on Jacobian norm and SelectiveInput Gradient Regularization (J-SIGR), which suggests the linearizedrobustness through Jacobian normalization and also regularizes theperturbation-based saliency maps to imitate the model's interpretablepredictions. As such, we achieve both the improved defense and highinterpretability of DNNs. Finally, we evaluate our method across differentarchitectures against powerful adversarial attacks. Experiments demonstratethat the proposed J-SIGR confers improved robustness against transferredadversarial attacks, and we also show that the predictions from the neuralnetwork are easy to interpret.

Jacobian Norm with Selective Input Gradient Regularization for Improved and Interpretable Adversarial Defense