Trustworthy Vision for Autonomous Vehicles: A Robust Logic-Infused Deep Learning Approach

Deep learning constitutes a fundamental pillar in the field of image recognition within autonomous vehicles (AVs), facilitating precise predictions based on unprocessed data. However, unlike human cognition, deep learning models are susceptible to adversarial attacks. This paper proposes a novel approach, termed the Robust Logic-infused Deep Learning (RLDL) Approach, designed for traffic sign recognition. RLDL employs Inductive Logic Programming (ILP) to derive logical rules from a combination of positive and negative examples. These rules are subsequently transformed into a matrix of logical constraints, allowing for the assessment of logical consistency in predictions. Then, this logical consistency is incorporated into the neural network through the loss function. This study explores the impact of integrating logical constraints into deep learning models on the reliability of vision tasks in AVs. Our experiments demonstrate that the proposed method substantially enhances the accuracy of recognising traffic signs under adversarial attacks.