Human Perception of AI Capabilities in Identifying Malicious Roadway Signs
Document Type
Abstract
Publication Date
2021
DOI
10.1037/tms0000077
Publication Title
TMS Proceedings 2021
Pages
1 pp.
Abstract
Artificial intelligence (AI) is widely utilized in various technologies such as Chatbots, voice assistants, and computer vision techniques for sensing roadway environments in automated vehicles (AVs). AI is required for AVs to perform and maintain safe roadway operations, but AI is never perfect and sometimes requires human assistance. Previously, we found that human drivers overestimated the AI's capability of identifying a maliciously manipulated stop sign. Manipulated signs interfered with the AI's recognition capabilities, and thus the safety of the system. In this study, we tested humans' understanding of different adversarial attacks to various road-sign images and their perception of the AI's capabilities in identifying these images. In addition to a baseline condition using original images and a control condition using pixel-shuffle images, there were to attack conditions including projected gradient descent (PGD) adversarial attacks and physical adversarial attacks. The PGD attack add subtle elements to the images that are not easily visible to human eyes, whereas the physical adversarial attacks are more obvious to human eyes; both attacks can lead the AI to misclassify the road signs. Our results showed that participants identified images in both attack conditions with similar accuracy to that of the original images. Participants rated the AI to be significantly less capable of identifying the original road signs and those with physical attacks than human eyes, although with small effect sized; there was no such difference for the PGD road signs. These results indicate participants' lack of understanding of the AI's vulnerability to PGD adversarial attacks.
Rights
© 2021 The Authors.
Published under a Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International (CC BY-NC-ND 4.0) License.
Original Publication Citation
Garcia, K. R., Xiao, Y., Mishler, S., Wang, C., Hu, B., & Chen, J. (2021, November 3). Human perception of AI capabilities in identifying malicious roadway signs. TMS Proceedings 2021 [Poster presentation]. Virtual. https://doi.org/10.1037/tms0000077
ORCID
0000-0001-9474-1907 (Garcia), 0000-0001-9104-1710 (Mishler)
Repository Citation
Garcia, Katherine R.; Xiao, Yanru; Mishler, Scott; Wang, Cong; Hu, Bin; and Chen, Jing, "Human Perception of AI Capabilities in Identifying Malicious Roadway Signs" (2021). Psychology Faculty Publications. 195.
https://digitalcommons.odu.edu/psychology_fac_pubs/195