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

Over the past decade, deep learning (DL) has been successfully applied tomany industrial domain-specific tasks. However, the current state-of-the-art DLsoftware still suffers from quality issues, which raises great concernespecially in the context of safety- and security-critical scenarios.Adversarial examples (AEs) represent a typical and important type of defectsneeded to be urgently addressed, on which a DL software makes incorrectdecisions. Such defects occur through either intentional attack orphysical-world noise perceived by input sensors, potentially hindering furtherindustry deployment. The intrinsic uncertainty nature of deep learningdecisions can be a fundamental reason for its incorrect behavior. Although sometesting, adversarial attack and defense techniques have been recently proposed,it still lacks a systematic study to uncover the relationship between AEs andDL uncertainty. In this paper, we conduct a large-scale study towards bridgingthis gap. We first investigate the capability of multiple uncertainty metricsin differentiating benign examples (BEs) and AEs, which enables to characterizethe uncertainty patterns of input data. Then, we identify and categorize theuncertainty patterns of BEs and AEs, and find that while BEs and AEs generatedby existing methods do follow common uncertainty patterns, some otheruncertainty patterns are largely missed. Based on this, we propose an automatedtesting technique to generate multiple types of uncommon AEs and BEs that arelargely missed by existing techniques. Our further evaluation reveals that theuncommon data generated by our method is hard to be defended by the existingdefense techniques with the average defense success rate reduced by 35\%. Ourresults call for attention and necessity to generate more diverse data forevaluating quality assurance solutions of DL software.

Towards Characterizing Adversarial Defects of Deep Learning Software from the Lens of Uncertainty