Interpretation Conflict in Helmet Recognition under Adversarial Attack

Abstract

Humans and Artificial Intelligence (AI) may have observation and interpretation conflicts in collaborative interaction. The adversarial samples make such conflicts more likely to occur in the field of image recognition. However, few studies have been seen combining the human-AI conflict and adversarial attack. This study presents the interpretation conflict due to adversarial samples in the helmet recognition task. A simulation also has been conducted to illustrate this problem. The results show that it should be prudent for the construction industry to land AI applications due to adversarial attacks on image recognition; the adversarial samples easily trigger interpretation conflicts, for example, the logo, graffiti, sticker, and text on helmets; lean construction should be propagated for the preconditions for AI applications.