Session: 03-16-01: Securing Advanced Manufacturing: Cybersecurity and Edge Computing for Industrial IoT

Paper Number: 164156

Evaluating the Impact of Cyberattacks on AI-Based Machine Vision Systems: A Case Study of Threaded Fasteners 

Machine vision systems integrated with Artificial Intelligence (AI) based technologies like Convolutional Neural Network (CNN) have shown promising results in classification and sorting applications. One such application is the classification of threaded fasteners like screws, bolts, nuts, washers, etc, during the assembly process to ensure that components are assembled securely. In previously published research, the authors have presented an image-based fastener classification model using a pre-trained CNN, specifically the Efficient-Net-b0. The model was integrated with an Explainable AI (XAI) system using Gradient-weighted Class Activation Maps (Grad-CAM) to explain the rationale for the classification results produced by Efficient-Net-b0. However, increasing penetration of the internet into manufacturing organizations introduces cybersecurity concerns that can impact the performance of AI-based machine vision systems. Cyberattacks such as data poisoning, and model evasion targeted towards AI systems can manipulate the input data or the model itself, leading to misclassification and potential operational failures. With an increase in machine vision systems being deployed in automated manufacturing, it is essential to safeguard the AI models against cyberattacks. Hence this research is motivated by the need to analyze the impact of one such class of cyberattacks, i.e. model evasion attack on an AI-based machine vision system. The focus of this research will be to analyze the impact of this attack on the accuracy and confidence of the model as applied to fastener classification.

When executed effectively, model evasion attacks can compromise a model's ability to classify images correctly, resulting in the model’s reduced confidence scores and/or inaccurate image classification. In this paper, we have utilized the Fast Gradient Sign Method (FGSM) to craft adversarial images of threaded fasteners, which are then used to perform model evasion attacks on the machine vision system described earlier. FGSM is particularly effective in white-box attacks, where the attacker has full knowledge of the model’s architecture and parameters. FGSM generates adversarial images of threaded fasteners quickly and efficiently by introducing subtle and nearly imperceptible modifications to the input images. It leverages the gradients of the model’s loss function to identify the optimal direction and magnitude of perturbations, ultimately deceiving the model into making incorrect predictions.

The overarching research questions answered in this study are: a) How does FGSM-generated adversarial images impact the accuracy and confidence of the classification results from the previously developed Efficient-Net-b0 based machine vision system? b) Is the XAI system implemented as part of the machine vision system capable of detecting FGSM-generated adversarial attacks? c) Can human-in-the-loop be a solution to mitigate the impact of FGSM-generated adversarial images? It was found that FGSM-generated adversarial images reduced the accuracy and confidence of the classification results. The XAI system detected adversarial images by displaying skewed heat maps that highlight the image regions that were the most influential in contributing to the model’s decision. Lastly, when humans are presented with FGSM-generated adversarial images, they are still able to identify the images with high accuracy and confidence. In conclusion, the paper evaluates the impact of a cyberattack on an AI-based machine vision system used in manufacturing, and also validates a potential solution to mitigate the impact of such an attack.

Presenting Author: Ankit Agarwal Clemson University

Presenting Author Biography: Ankit Agarwal is a Faculty and Postdoctoral Researcher in the Automotive Engineering department at Clemson University, SC. He teaches and conducts research in CAD/CAM, modeling of manufacturing processes, tool wear analysis, and application of data science in manufacturing. He is currently working on an NSF-funded project on analyzing tool wear in Nickel-based superalloys.
Prior to joining Clemson University, he did his Ph.D. with a specialization in Manufacturing Engineering from the Indian Institute of Technology Jodhpur, Rajasthan, India, in 2021. He worked on modeling and control of geometric tolerance in end milling of thin-walled components.

Authors:

Vijayanth Tummala Ball State University
Ankit Agarwal Clemson University
Amaninder Singh Gill Centralia College
Seung-Jin Lee University of Washinton-Tacoma
Laine Mears Clemson University