Abstract
Multi-agent perception systems, such as Vehicle-to-Vehicle (V2V) communication, enhance autonomous driving by allowing vehicles to share sensor data. However, this collaboration depends heavily on accurate localization for spatial alignment. This paper presents AdvGPS, a novel adversarial attack that demonstrates how stealthy, strategically crafted GPS errors can disrupt cooperative perception. By nudging a vehicle's reported pose within the bounds of natural GPS noise, AdvGPS causes state-of-the-art models to miss real objects or detect "ghost" objects, significantly degrading performance.
Introduction: The Vulnerability of Cooperative Perception
In autonomous driving, "two heads are better than one." V2V communication enables vehicles to share LiDAR point clouds, allowing them to see through obstructions or around corners. For an "Ego" vehicle to utilize data from a Connected and Autonomous Vehicle (CAV), it must project that data into its own coordinate system using the GPS pose (position and heading) of both agents.
While GPS is known to be imprecise, most systems treat errors as random noise. AdvGPS challenges this assumption by asking: Can a malicious actor craft specific, non-random GPS errors that appear natural but are calculated to break the perception system?
Methodology: Crafting Stealthy Perturbations
AdvGPS is a black-box adversarial attack that modifies the six parameters of a vehicle’s pose ($x, y, z$ and roll, pitch, yaw). To ensure effectiveness and transferability across different AI architectures, the attack optimizes the GPS signal using three "statistically sensitive natural discrepancies":
- Appearance-based Discrepancy ($D_{app}$): This maximizes the Mean Squared Error between the original and "attacked" projected point clouds, targeting raw spatial misalignment.
- Distribution-based Discrepancy ($D_{dist}$): Using Maximum Mean Discrepancy (MMD), this ensures that features extracted by the neural network differ fundamentally from the original ones, disrupting the high-level "understanding" of the scene.
- Task-aware Discrepancy ($D_{task}$): A surrogate model calculates a loss function based on object detection performance, pushing the system toward classification and regression errors.
A critical constraint of AdvGPS is stealthiness. Perturbations are kept within real-world GPS error bounds (e.g., ~1.1 meters for horizontal position), making the attack difficult to distinguish from routine signal degradation.
Experimental Evaluation
The researchers evaluated AdvGPS using the OPV2V dataset, a benchmark for V2V perception simulated in CARLA and OpenCDA. The attack was tested against five state-of-the-art fusion methods: AttFuse, F-cooper, V2VAM, V2X-ViT, and CoBEVT.
Key Results
The quantitative impact of AdvGPS is severe. Using Average Precision (AP) at an Intersection over Union (IoU) of 0.5:
- CoBEVT: Performance dropped from 0.904 (no attack) to 0.403 under the AdvGPS black-box attack.
- V2X-ViT: Dropped from 0.928 to 0.434.
- Comparison: AdvGPS consistently outperformed baseline attacks. For F-Cooper, AdvGPS reduced AP to 0.298, whereas a standard Position Altering Attack (PAA) only reached 0.524.
Discussion and Insights
- "Worse Than Nothing": AdvGPS can reduce accuracy to levels below what a single car would achieve without any shared data ("No Fusion"). A compromised agent thus becomes a liability.
- Spatial Sensitivity: Positional coordinates ($x, y$) are more sensitive to attack than heading. Small shifts in position cause massive misalignments in high-resolution LiDAR data.
- Transferability: By optimizing for general statistical discrepancies, attacks generated against a simple surrogate model remain highly effective against complex, "unseen" models like V2X-ViT.
- Stealth as a Weapon: Because the point cloud shifts are minimal (approx. 1 meter), they appear as normal noise to human observers or simple outlier detectors, yet they are mathematically tuned to confuse fusion AI.
Limitations and Future Work
This study was conducted in simulation; real-world "urban canyon" GPS noise may present different challenges. Furthermore, while the paper highlights a critical security hole, it does not propose a specific defense mechanism. The authors suggest that robust fusion and multi-sensor cross-checking are vital areas for future research to secure multi-agent systems.
Summary updated