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 network image classifiers are known to be susceptible not only toadversarial examples created for them but even those created for others. Thisphenomenon poses a potential security risk in various black-box systems relyingon image classifiers. The reason behind such transferability of adversarialexamples is not yet fully understood and many studies have proposed trainingmethods to obtain classifiers with low transferability. In this study, weaddress this problem from a novel perspective through investigating thecontribution of the network architecture to transferability. Specifically, wepropose an architecture searching framework that employs neuroevolution toevolve network architectures and the gradient misalignment loss to encouragenetworks to converge into dissimilar functions after training. Our experimentsshow that the proposed framework successfully discovers architectures thatreduce transferability from four standard networks including ResNet and VGG,while maintaining a good accuracy on unperturbed images. In addition, theevolved networks trained with gradient misalignment exhibit significantly lowertransferability compared to standard networks trained with gradientmisalignment, which indicates that the network architecture plays an importantrole in reducing transferability. This study demonstrates that designing orexploring proper network architectures is a promising approach to tackle thetransferability issue and train adversarially robust image classifiers.

Evolving Architectures with Gradient Misalignment toward Low Adversarial Transferability