이승열, 하재철,

Vol. 35, No. 2, pp. 347-358, 4월. 2025
10.13089/JKIISC.2025.35.2.347, Full Text:
Keywords: autonomous vehicles, Imitation Learning, adversarial attacks
Abstract

Recently, artificial intelligence technology has been commercially utilized in various advanced industrial fields, such as autonomous driving, home networks, and smart factories. In particular, significant research has been conducted on ADAS(Advanced Driver Assistance Systems) in autonomous vehicles. However, it has been found to be vulnerable to adversarial attacks. In this paper, we constructed a deep learning model based on imitation learning to predict the steering angle and throttle of autonomous vehicles and applied it to the CARLA simulation to perform adversarial attacks. The performance of the model was evaluated using MSE(Mean Squared Error) and MAE(Mean Absolute Error). Furthermore, this paper proposes the GRM(Gradient Removal Module), an improved version of GSI(Gradient Sign Inversion), as a defense mechanism against adversarial attacks. The experimental results demonstrate that the proposed method, when applied to imitation learning-based deep learning models and evaluated using MSE and MAE, achieves superior performance compared to existing defense techniques, with an MSE of 0.0022 and an MAE of 0.0198.

Statistics
Cite this article