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

Adversarial training based on the minimax formulation is necessary forobtaining adversarial robustness of trained models. However, it is conservativeor even pessimistic so that it sometimes hurts the natural generalization. Inthis paper, we raise a fundamental question---do we have to trade off naturalgeneralization for adversarial robustness? We argue that adversarial trainingis to employ confident adversarial data for updating the current model. Wepropose a novel approach of friendly adversarial training (FAT): rather thanemploying most adversarial data maximizing the loss, we search for leastadversarial (i.e., friendly adversarial) data minimizing the loss, among theadversarial data that are confidently misclassified. Our novel formulation iseasy to implement by just stopping the most adversarial data searchingalgorithms such as PGD (projected gradient descent) early, which we callearly-stopped PGD. Theoretically, FAT is justified by an upper bound of theadversarial risk. Empirically, early-stopped PGD allows us to answer theearlier question negatively---adversarial robustness can indeed be achievedwithout compromising the natural generalization.

Attacks Which Do Not Kill Training Make Adversarial Learning Stronger