Securing Autonomous Systems: Taxonomy, Challenges, and Defense Mechanisms Against Adversarial Threats

Autonomous Systems (AS) have become the future of many industries, yet they face escalating adversarial threats. These attacks exploit vulnerabilities in deep neural networks (DNNs), leading to severe malfunctions and safety concerns. This paper introduces a comprehensive taxonomy of adversarial threats tailored to AS, encompassing both digital and physical domains. Unlike prior work, we emphasize underexplored areas such as real-world physical attacks and their implications for system-level security. To address the limitations of current methodologies, we propose novel evaluation frameworks designed to assess lifecycle-wide resilience, integrating system-level reliability, real-world adaptability, and proactive defense mechanisms. Our contributions provide actionable insights and a roadmap for advancing the security of next-generation autonomous technologies. State-of-the-art defense mechanisms are systematically classified, evaluated, and rated based on their robustness and practicality for real-world integration by identifying the key requirements for robust AS, emphasizing the need for adaptive defenses tailored to the dynamic and evolving threat landscape. Open issues, including gaps in benchmarks, scalability, and the lack of life-cycle-wide security frameworks, are discussed alongside actionable research directions.