Defenses for Visual Recognition in Connected Autonomous Vehicles: A Review

Connected Autonomous Vehicles (CAVs) promise enhanced safety, efficiency, and accessibility in transportation by operating autonomously while communicating with road infrastructure and other vehicles. The development of CAV technology involves advanced vision-based classification systems, utilizing deep neural networks (DNNs) to perceive and interpret sensor data for navigation. However, the increased reliance on automation technology makes CAVs susceptible to adversarial attacks—malicious inputs designed to disrupt model decision-making, potentially leading to safety and reliability hazards. Traditional and modern classification algorithms have shown susceptibility to various forms of adversarial perturbations, which can lead to potential erroneous decisions in real-world scenarios. This review explores the transformative potential and challenges posed by CAVs and associated visual recognition technologies, particularly DNNs and compressive sensing (CS) techniques that can be used to defend against adversarial attacks on CAVs. This review examines various types of adversarial attacks, including the Fast Gradient Sign Method (FGSM), Carlini & Wagner (CW), and sticker attacks, which exploit CAV visual recognition system vulnerabilities. These attacks pose significant risks to CAV decision-making, emphasizing the critical need for robust defense mechanisms. Several defensive approaches are analyzed, including compressive sensing, incremental training, and classifier-aware training, which each offering unique advantages in strengthening CAV systems against adversarial threats. This review highlights the need for adaptive defense strategies specifically designed for CAV visual recognition systems, proposing the integration of these techniques as a critical step toward securing the future of autonomous transportation.