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 Tools annual

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 Tools monthly

Global: Zendy Plus monthly

Presents the access of premium content as premium feature

Premium Content

Global: Zendy Plus annual

Global: Zendy Free Plan

Human Activity Recognition (HAR) has been employed in a wide range ofapplications, e.g. self-driving cars, where safety and lives are at stake.Recently, the robustness of existing skeleton-based HAR methods has beenquestioned due to their vulnerability to adversarial attacks, which causesconcerns considering the scale of the implication. However, the proposedattacks require the full-knowledge of the attacked classifier, which is overlyrestrictive. In this paper, we show such threats indeed exist, even when theattacker only has access to the input/output of the model. To this end, wepropose the very first black-box adversarial attack approach in skeleton-basedHAR called BASAR. BASAR explores the interplay between the classificationboundary and the natural motion manifold. To our best knowledge, this is thefirst time data manifold is introduced in adversarial attacks on time series.Via BASAR, we find on-manifold adversarial samples are extremely deceitful andrather common in skeletal motions, in contrast to the common belief thatadversarial samples only exist off-manifold. Through exhaustive evaluation, weshow that BASAR can deliver successful attacks across classifiers, datasets,and attack modes. By attack, BASAR helps identify the potential causes of themodel vulnerability and provides insights on possible improvements. Finally, tomitigate the newly identified threat, we propose a new adversarial trainingapproach by leveraging the sophisticated distributions of on/off-manifoldadversarial samples, called mixed manifold-based adversarial training (MMAT).MMAT can successfully help defend against adversarial attacks withoutcompromising classification accuracy.

Understanding the Vulnerability of Skeleton-based Human Activity Recognition via Black-box Attack