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

Fast adversarial training (FAT) is an efficient method to improve robustness.However, the original FAT suffers from catastrophic overfitting, whichdramatically and suddenly reduces robustness after a few training epochs.Although various FAT variants have been proposed to prevent overfitting, theyrequire high training costs. In this paper, we investigate the relationshipbetween adversarial example quality and catastrophic overfitting by comparingthe training processes of standard adversarial training and FAT. We find thatcatastrophic overfitting occurs when the attack success rate of adversarialexamples becomes worse. Based on this observation, we propose a positiveprior-guided adversarial initialization to prevent overfitting by improvingadversarial example quality without extra training costs. This initializationis generated by using high-quality adversarial perturbations from thehistorical training process. We provide theoretical analysis for the proposedinitialization and propose a prior-guided regularization method that boosts thesmoothness of the loss function. Additionally, we design a prior-guidedensemble FAT method that averages the different model weights of historicalmodels using different decay rates. Our proposed method, called FGSM-PGK,assembles the prior-guided knowledge, i.e., the prior-guided initialization andmodel weights, acquired during the historical training process. Evaluations offour datasets demonstrate the superiority of the proposed method.

Improving Fast Adversarial Training with Prior-Guided Knowledge