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 has been proven to be a powerful regularization methodto improve the generalization of models. However, current adversarial trainingmethods only attack the original input sample or the embedding vectors, andtheir attacks lack coverage and diversity. To further enhance the breadth anddepth of attack, we propose a novel masked weight adversarial training methodcalled DropAttack, which enhances generalization of model by addingintentionally worst-case adversarial perturbations to both the input and hiddenlayers in different dimensions and minimize the adversarial risks generated byeach layer. DropAttack is a general technique and can be adopt to a widevariety of neural networks with different architectures. To validate theeffectiveness of the proposed method, we used five public datasets in thefields of natural language processing (NLP) and computer vision (CV) forexperimental evaluating. We compare the proposed method with other adversarialtraining methods and regularization methods, and our method achievesstate-of-the-art on all datasets. In addition, Dropattack can achieve the sameperformance when it use only a half training data compared to other standardtraining method. Theoretical analysis reveals that DropAttack can performgradient regularization at random on some of the input and wight parameters ofthe model. Further visualization experiments show that DropAttack can push theminimum risk of the model to a lower and flatter loss landscapes. Our sourcecode is publicly available on https://github.com/nishiwen1214/DropAttack.

DropAttack: A Masked Weight Adversarial Training Method to Improve Generalization of Neural Networks