Abstract: In the current landscape of autonomous vehicle (AV) safety and security research, there are multiple isolated problems being tackled by the community at large. Due to the lack of common evaluation criteria, several important research questions are at odds with one another. For instance, while much research has been conducted on physical attacks deceiving AV perception systems, there is often inadequate investigations on working defenses and on the downstream effects of safe vehicle control. This paper provides a thorough description of the current state of AV safety and security research. We provide individual sections for the primary research questions that concern this research area, including AV surveillance, sensor system reliability, security of the AV stack, algorithmic robustness, and safe environment interaction. We wrap up the paper with a discussion of the issues that concern the interactions of these separate problems. At the conclusion of each section, we propose future research questions that still lack conclusive answers. This position article will serve as an entry point to novice and veteran researchers seeking to partake in this research domain.
Introduction: The Fragmentation of AV Research
The current state of autonomous vehicle (AV) research is dangerously fragmented. Security and safety are often treated as separate domains, leading to a "cat-and-mouse game" played in silos. Security researchers may prove they can trick an AV’s perception system with an adversarial patch, yet they rarely investigate if that trick actually causes a safety failure, such as a crash. Conversely, safety engineers often design robust control systems while assuming the underlying sensor data and middleware are unhackable. This paper argues that the traditional "Sense-Plan-Act" model is insufficient for modern security analysis and proposes a holistic, end-to-end view of the AV pipeline.
The 5-Stage AV Pipeline
To identify where vulnerabilities hide and how they propagate, we propose a granular 5-stage pipeline. This framework moves beyond isolated component testing to evaluate how a breach in one stage impacts the entire system's functional safety.
- Environment State: The physical world and the privacy of those within it.
- Sensor System: The hardware (LiDAR, cameras) and raw data processing.
- Compute Environment: The "hidden" layer of middleware, such as the Robot Operating System (ROS) or ArduPilot, and hardware resources.
- Algorithm: The AI-driven tracking, navigation logic, and control systems.
- Environment Interaction: The physical movement of the vehicle and the resulting legal and regulatory framework.
Key Insights and Vulnerabilities
The Middleware Blind Spot
Much of the industry relies on ROS or ROS2. These systems often have significant security vulnerabilities, such as a lack of authentication, which could allow an attacker to hijack internal communications. These middleware risks are frequently ignored in high-level safety discussions but represent a critical point of failure.
Downstream Effects vs. Accuracy
Traditional AI research focuses on classification accuracy. However, in an AV context, a misclassification only matters if it impacts vehicle control. If a tracking system filters out a momentary error, the attack has zero safety impact. We advocate for metrics that measure downstream effects on vehicle behavior rather than isolated component performance.
Timing and Resource Exhaustion
In safety-critical systems, timing is everything. A perfectly accurate detection that arrives 100ms too late is a failure. Security attacks targeting timing or energy consumption—such as forcing a GPU to overwork to induce latency—are just as dangerous as adversarial images.
Privacy as a Safety Risk
The push for more data to improve safety leads to mass surveillance concerns. If the public fears AVs due to privacy risks, they may adopt "anti-surveillance" measures (like adversarial clothing) that inadvertently make sensors less reliable, creating new safety hazards.
Future Research Recommendations
To bridge the gap between security and safety, we propose three major research directions:
- End-to-End Testing: Move away from testing components in isolation. Universal metrics are needed to weigh how an attack on a sensor propagates through the middleware to the steering wheel.
- Cooperative Perception: Utilize Vehicle-to-Everything (V2X) communication to cross-check sensor data across multiple vehicles, making it significantly harder for an attacker to spoof a single vehicle's reality.
- Software Licensing: Governments should consider licensing specific versions of AV software for specific environments, ensuring provable functional safety before deployment.
Conclusion
Achieving the safety and security of autonomous vehicles requires a shift from probabilistic defenses to provably secure architectures. By viewing the AV pipeline as an interconnected whole, researchers can better address the "sim-to-real gap" and the unpredictable nature of human behavior in the physical environment.
Summary updated