Is Your Autonomous Vehicle Safe? Understanding the Threat of Electromagnetic Signal Injection Attacks on Traffic Scene Perception

Wenhao Liao, Sineng Yan, Youqian Zhang, Xinwei Zhai, Yuanyuan Wang, Eugene Yujun FuAAAI 2025 · DOI

Introduction

Autonomous vehicles (AVs) rely heavily on camera-based perception systems to interpret their surroundings. While digital adversarial attacks—such as physical patches on road signs—are well-documented, a more insidious threat exists at the hardware level: Electromagnetic Signal Injection Attacks (ESIA). By placing a small device near a vehicle's camera, an attacker can inject electromagnetic interference directly into the circuitry, causing horizontal color strip distortions that bypass traditional digital filters.

Overview of the ESIA attack pipeline on an autonomous vehicle perception system.

This research addresses the critical gap in understanding how these hardware-level distortions impact the "brains" of autonomous systems. Because collecting real-world attack data is dangerous and expensive, the authors developed a high-fidelity simulation method to evaluate AI model resilience across thousands of driving scenarios.

Methodology: Simulating Hardware Failure

To accurately model ESIA, the researchers moved beyond simple noise addition. They analyzed how electromagnetic signals cause packet loss and incorrect color reconstruction during the camera's image processing phase.

The resulting simulation method models the swapping of pixel values between Red, Green, and Blue channels within specific "Impacted Rows." For example, if a row index is even, the red channel might take values from the green channel of the next pixel, effectively creating the characteristic color strips seen in real-world hardware interference.

The simulation process for ESIA, demonstrating how pixel values are swapped between color channels to create adversarial patterns.

Statistical validation using t-tests on Mean Average Precision (mAP) confirmed that there was no significant performance difference between AI models processing these simulated images versus real-world attack data, establishing the simulation as a reliable proxy for hardware-level threats.

Experimental Setup and Evaluation

The study utilized the BDD100k dataset, selecting 10,000 images to create an ESIA Simulation Dataset categorized by severity (Mild, Moderate, Severe) and environmental conditions (Weather, Scene, Time of Day). The researchers evaluated 11 object detection architectures, including YOLOv3 and DETR, alongside multi-task models like HybridNets and YOLOP that handle both detection and drivable area segmentation.

Key Insights

The "Simplicity" Paradox

One of the most striking findings was that AI models performed significantly worse under attack in "simple" environments, such as highways, compared to complex city streets. In highway scenes, detection performance dropped by an average of 70.89% under severe attack. The authors hypothesize that in cluttered city environments, models are forced to focus on finer details, providing a slight inherent resilience to the coarse horizontal distortions of ESIA.

Task-Specific Vulnerability

The research revealed that object detection (identifying cars and pedestrians) is far more fragile than drivable area segmentation. Under severe attack, detection accuracy (mAP) plummeted by 60-70%, while segmentation (mIoU) dropped by only 11-14%. However, even small segmentation errors can be fatal; qualitative results showed ESIA could trick a model into identifying a reverse lane as a valid drivable path.

Attention Hijacking

Using Grad-CAM visualizations, the study demonstrated that ESIA causes the AI's "attention" to shift away from actual obstacles and toward the edges of the color strips. This effectively blinds the vehicle to real-world hazards, as the neural network prioritizes the high-contrast boundaries of the injected interference over actual traffic objects.

Conclusion and Future Work

ESIA represents a significant physical threat to AV safety, capable of causing catastrophic perception failures. While some models like A-YOLOM and YOLOP showed relative stability in specific tasks, the overall vulnerability of current architectures is high.

Future research will focus on multi-sensor fusion resilience, investigating whether LiDAR and Radar data can mitigate camera-specific failures. Additionally, the authors plan to move from "open-loop" image analysis to "closed-loop" simulations to observe how these perception errors translate into real-time steering and braking decisions.

Summary updated

BibTeX
@inproceedings{liao2025is,
  author = {Wenhao Liao and Sineng Yan and Youqian Zhang and Xinwei Zhai and Yuanyuan Wang and Eugene Yujun Fu},
  title = {{Is Your Autonomous Vehicle Safe? Understanding the Threat of Electromagnetic Signal Injection Attacks on Traffic Scene Perception}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i26.34958},
}
Figure