Evaluating the Vulnerability of YOLOv5 to Adversarial Attacks for Enhanced Cybersecurity in MASS
Changui Lee, Seojeong Lee
Abstract
The development of artificial intelligence (AI) technologies, such as machine learning algorithms, computer vision systems, and sensors, has allowed maritime autonomous surface ships (MASS) to navigate, detect and avoid obstacles, and make real-time decisions based on their environment. Despite the benefits of AI in MASS, its potential security threats must be considered. An adversarial attack is a security threat that involves manipulating the training data of a model to compromise its accuracy and reliability. This study focuses on security threats faced by a deep neural network-based object classification algorithm, particularly you only look once version 5 (YOLOv5), which is a model used for object classification. We performed transfer learning on YOLOv5 and tested various adversarial attack methods. We conducted experiments using four types of adversarial attack methods and parameter changes to determine the attacks that could be detrimental to YOLOv5. Through this study, we aim to raise awareness of the vulnerability of AI algorithms for object detection to adversarial attacks and emphasize the need for efforts to overcome them; these efforts can contribute to safe navigation in MASS.