Adversarial Attacks in Autonomous Vehicles: A Comprehensive Review of Threats and Defence Strategies

Abstract

Autonomous Vehicles (AVs) are becoming a core component of modern transportation systems, utilising a wide range of sensor data and Machine Learning (ML) models to perceive and interpret real-world driving environments. Adversarial attacks can deliberately manipulate sensor data and models, posing serious threats to the performance, safety, security, robustness, deployability and reliability of AVs. Safeguarding key requirements of AV technologies has emerged as a critical and timely research challenge. This article reviews existing adversarial threat models and attack techniques, covering their impact on key AV components, including perception systems, communication modules, decision-making, and planning processes. Adversarial attacks are classified into a five-dimensional taxonomy based on the attacker’s knowledge level, target, objective, domain, and specificity. A broad range of defence strategies is examined and organised into three levels - input-level, model-level, and system-level - reflecting the stage at which protection is applied. The effectiveness of attacks and defences is evaluated with several metrics. Metrics, including Attack Success Rate (ASR), transferability, and query count, are considered for attack performance evaluation. For defence methods, performance is measured using accuracy, precision, Structural Similarity Index Measure (SSIM), and detection rate. Several limitations are observed in current evaluation practices. In particular, the dependence on general-purpose benchmarks, limited diverse attack scenarios, and the lack of reports on computational complexity reduce the practical relevance of reported results. Therefore, these limitations reduce the reliability of real-world deployment. In addition, exploring the trade-offs between robust security, model accuracy, and feasibility in deployment remains insufficient. In this review, we discuss these issues in more detail and highlight directions for future research.