An illumination based backdoor attack against crack detection systems in laser beam welding

Publication date
2025-05-27
Document type
Konferenzbeitrag
Author
Huo, Wenjie
Schmies, Lennart
Gumenyuk, Andrey
Rethmeier, Michael
Wolter, Katinka
Organisational unit
Mathematics and Computer Science, Free University Berlin
BAM - Bundesanstalt fĂĽr Materialforschung und -prĂĽfung, Fachbereich SchweiĂźtechnische Fertigungsverfahren, Berlin
DOI
URI
Conference
Publisher
Universitätsbibliothek der HSU/UniBw H
Book title
Machine learning for cyber physical systems : proceedings of the conference ML4CPS 2025
First page
12
Last page
21
Is part of
Part of the university bibliography
Nein
Files
Language
English
Keyword
Welding crack detection
Backdoor attack
Deep neural networks
System security
Abstract
Deep neural networks (DNNs) have been wildly used in engineering and have achieved state-of-the-art performance in prediction and measurement tasks. A solidification crack is a serious fault during laser beam welding and it has been proven to be successfully detected using DNNs. Recently, research on the security of DNNs is receiving increasing attention because it is necessary to explore the reliability of DNNs to avoid potential security risks. The backdoor attack is a serious threat, where attackers aim to inject an inconspicuous pattern referred to as trigger into a small portion of training data, resulting in incorrect predictions in the reference phase whenever the input contains the trigger. In this work, we first generate experimental data containing actual cracks in the welding laboratory for training a crack detection model. Then, targeting this scenario, we design a new type of backdoor attack to induce the model to predict the crack as a normal state. Considering the stealthiness of the attack, a common phenomenon during the welding process, illumination, is used as the backdoor trigger. Experimental results demonstrate that the proposed method can successfully attack the crack detection system and achieve over 90% attack success rate on the test set.
Version
Published version
Access right on openHSU
Open access