WIP: End-to-End Analysis of Adversarial Attacks to Automated Lane Centering Systems

Hengyi Liang, Ruochen Jiao, Takami Sato, Junjie Shen, Qi Alfred Chen, Qi Zhu2021 · DOI

Abstract

Machine learning techniques, particularly those based on deep neural networks (DNNs), are widely adopted in the development of advanced driver-assistance systems (ADAS) and autonomous vehicles. While providing significant improvement over traditional methods in average performance, the usage of DNNs also presents great challenges to system safety, especially given the uncertainty of the surrounding environment, the disturbance to system operations, and the current lack of methodologies for predicting DNN behavior. In particular, adversarial attacks to the sensing input may cause errors in systems' perception of the environment and lead to system failure. However, existing works mainly focus on analyzing the impact of such attacks on the sensing and perception results and designing mitigation strategies accordingly. We argue that as system safety is ultimately determined by the actions it takes, it is essential to take an end-to-end approach and address adversarial attacks with the consideration of the entire ADAS or autonomous driving pipeline, from sensing and perception to planing, navigation and control. In this paper, we present our initial findings in quantitatively analyzing the impact of a type of adversarial attack (that leverages road patch) on system planning and control, and discuss some of the possible directions to systematically address such attack with an end-to-end view.

Introduction

Most adversarial research in the autonomous vehicle (AV) space focuses on fooling a car's vision—for instance, making a stop sign appear as a speed limit sign. However, in a real-world driving scenario, a perception error is only dangerous if it translates into a hazardous physical action. This "Work-in-Progress" (WIP) research shifts the focus from isolated perception errors to the "end-to-end" impact on steering and control. By analyzing the full pipeline of an Automated Lane Centering (ALC) system, we demonstrate how physical adversarial attacks propagate through perception and planning to cause dangerous maneuvers.

System Architecture and Methodology

To conduct a high-fidelity analysis, we utilized a production-grade software stack, OpenPilot, integrated with the LGSVL simulator. This setup provides a closed-loop environment where the vehicle's steering actions directly influence the sensor input of subsequent frames.

The end-to-end pipeline of the ALC system, showing the flow from perception to planning and control.

The system architecture is divided into three primary modules:

  1. Perception Module: Uses a combination of Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) to detect lane lines and predict paths based on visual input.
  2. Planning Module: Determines the "desired path" by calculating a weighted average of predicted paths and detected lane lines.
  3. Control Module: Employs Model Predictive Control (MPC) to translate the desired path into physical steering angles and speed adjustments.

The attack vector used is a "dirty road patch"—a physical-world perturbation optimized to appear as harmless road grime to human eyes but mathematically designed to mislead the DNNs responsible for lane detection.

Key Insights: The Confidence Canary and the Planning Trap

Our analysis revealed two critical vulnerabilities in how ALC systems handle uncertainty.

1. Confidence as an Early Warning

A significant finding is that the perception module’s confidence scores for lane detection drop sharply approximately 30 meters before the vehicle physically reaches the adversarial patch. This drop occurs before the vehicle begins to swerve, suggesting that internal confidence metrics could serve as a "canary in the coal mine" for detecting adversarial interference.

Confidence scores for lane detection drop significantly before the vehicle reaches the adversarial patch, acting as an early warning signal.

2. The Planning Trap

OpenPilot calculates its final trajectory as a weighted average. When the adversarial patch causes confidence in the lane lines to drop, the planning module automatically shifts more weight toward its "predicted driving path." Because the attack is specifically designed to bias this internal prediction, the system's own uncertainty-handling logic inadvertently facilitates the success of the attack, leading the vehicle toward the edge of the road or into oncoming traffic.

Results and Impact

The severity of the attack is tied to the Perturbation Area Ratio (PAR)—the percentage of patch pixels modified.

  • Total Failure: At a PAR of 30% to 50%, the vehicle consistently failed, driving off the road or crossing lane boundaries in less than one second.
  • Subtle Instability: Even at a low PAR of 10%, where the vehicle remained within its lane, the attack significantly degraded steering smoothness and overall vehicle stability.

Conclusion and Future Work

This study highlights that AV safety is not merely a matter of high-accuracy vision; it is about how planning and control modules interpret ambiguous data. Future mitigation strategies should focus on:

  • Sensor Fusion: Cross-referencing camera data with LiDAR or GPS when perception confidence spikes.
  • Temporal Logic: Enhancing RNNs to leverage historical data and ignore sudden, localized noise from road patches.
  • Holistic Adaptation: Developing frameworks where the vehicle automatically adopts a conservative driving posture or returns control to the human driver when uncertainty exceeds a safe threshold.

We plan to extend this methodology to other ADAS features, such as Adaptive Cruise Control, to build a more comprehensive understanding of end-to-end system vulnerabilities.

Summary updated

Abstract

Machine learning techniques, particularly those based on deep neural networks (DNNs), are widely adopted in the development of advanced driver-assistance systems (ADAS) and autonomous vehicles. While providing significant improvement over traditional methods in average performance, the usage of DNNs also presents great challenges to system safety, especially given the uncertainty of the surrounding environment, the disturbance to system operations, and the current lack of methodologies for predicting DNN behavior. In particular, adversarial attacks to the sensing input may cause errors in systems' perception of the environment and lead to system failure. However, existing works mainly focus on analyzing the impact of such attacks on the sensing and perception results and designing mitigation strategies accordingly. We argue that as system safety is ultimately determined by the actions it takes, it is essential to take an end-to-end approach and address adversarial attacks with the consideration of the entire ADAS or autonomous driving pipeline, from sensing and perception to planing, navigation and control. In this paper, we present our initial findings in quantitatively analyzing the impact of a type of adversarial attack (that leverages road patch) on system planning and control, and discuss some of the possible directions to systematically address such attack with an end-to-end view.
BibTeX
@misc{liang2021wip,
  author = {Hengyi Liang and Ruochen Jiao and Takami Sato and Junjie Shen and Qi Alfred Chen and Qi Zhu},
  title = {{WIP: End-to-End Analysis of Adversarial Attacks to Automated Lane Centering Systems}},
  year = {2021},
  doi = {10.14722/autosec.2021.23034},
}
Figure