Universal Adversarial Attacks for Visual Odometry Systems

Visual Odometry (VO) has gained significant attention as a critical technology with broad applications in autonomous navigation, augmented reality, and related domains. However, recent research indicates that VO systems are susceptible to adversarial attacks, leading to compromised accuracy and potential system failure. Traditional adversarial attack algorithms, primarily relying on random perturbations or objective function minimization, are not suitable for VO algorithms. In this paper, we present a novel and general adversarial attack algorithm specifically designed for targeting the yaw and translation components of visual odometry, while increasing the Euclidean distance between adjacent frames. Through a comprehensive analysis of VO algorithm characteristics, we propose an effective approach to disrupt VO system operation. Extensive experimental results demonstrate that the proposed attack algorithm significantly reduces the localization accuracy of VO algorithms while exhibiting robustness and generality. The findings of this research contribute to enhancing the security and stability of deep learning-based visual odometry algorithms, providing valuable insights and guidance for practical applications.