Privacy-Aware Sharing of Raw Spatial Sensor Data for Cooperative Perception

Bangya Liu, Chengpo Yan, Chenghao Jiang, Suman Banerjee, Akarsh PrabhakaraarXiv 2025 · DOI · arXiv

Introduction

Cooperative perception is a transformative technology for autonomous vehicles (AVs), allowing them to share sensor data to "see" beyond their immediate line-of-sight. While sharing processed data like bounding boxes is common, the exchange of raw spatial data (Lidar, Camera, Radar) offers the highest potential for perception accuracy, especially in complex or adversarial environments. However, sharing raw data introduces significant risks: location leakage and intellectual property (IP) theft. This paper presents SHARP (SHAring Raw spatial sensor data Privately), a framework designed to enable high-utility raw data sharing while safeguarding privacy and proprietary technology.

The Privacy and IP Challenge

The adoption of raw data sharing is hindered by two primary concerns. First, Location Leakage: raw sensor streams are rich enough to allow an observer to perform visual localization, pinpointing a vehicle's exact GPS-level pose even if pseudonyms are used. Second, Intellectual Property (IP) Theft: raw data reveals a manufacturer's specific sensor configurations, noise profiles, and low-level processing techniques—essentially their "secret sauce."

Overview of cooperative perception and the sensor data stack highlighting privacy and IP concerns.

The SHARP Framework

SHARP addresses these challenges through two innovative mechanisms:

1. Novel View Synthesis (NVS) for Location Privacy

To prevent location tracking, SHARP decouples the shared data from the vehicle's actual physical coordinates. Instead of transmitting the direct sensor feed, the vehicle uses a 3D vision foundation model, such as Visual Geometry Grounded Transformers (VGGT), to reconstruct the local 3D scene. It then renders a synthetic raw data stream from a randomized novel viewpoint near the vehicle. This "spoofed" data remains highly useful for the receiver's perception tasks but obscures the sender's true location.

SHARP's Novel View Synthesis (NVS) pipeline for generating privacy-preserving sensor data from randomized viewpoints.

2. Open-Stack Swapping for IP Protection

To protect proprietary sensor designs, SHARP employs a dual-stack architecture. While the vehicle uses its high-performance Proprietary Stack for internal navigation, it periodically switches to a standardized, hardware-agnostic Open Stack to generate data for sharing. This open stack utilizes public-domain algorithms for denoising and stabilization, effectively masking the unique hardware advantages of the sender.

Key Results and Insights

The researchers evaluated SHARP using the OPV2V dataset, the CARLA simulator, and nuScenes. Their findings underscore the necessity and effectiveness of the framework:

  • Vulnerability of Raw Data: Without obfuscation, an adversary can estimate a vehicle's pose with an accuracy (RMSE) of 0.69 meters using foundation models.
  • Privacy Gains: SHARP's randomized viewpoints increased the location estimation error to over 45 meters, successfully thwarting tracking attempts.
  • Utility Preservation: Despite the synthetic nature of the shared views, receiving vehicles still achieved significantly lower depth estimation errors compared to operating without cooperative data.
  • Real-time Performance: Modern models like VGGT can perform inference in approximately 89ms, enabling 10Hz real-time cooperative perception.

Conclusion and Future Directions

SHARP demonstrates that privacy and high-fidelity cooperative perception are not mutually exclusive. Future work will focus on improving 3D reconstruction in highly dynamic environments with fast-moving pedestrians and cars, ensuring occlusion-aware viewpoint selection, and developing techniques to mitigate potential adversarial rendering detection. Additionally, industry-wide standardization of the "Open Stack" remains a critical step for widespread deployment.

Summary updated

BibTeX
@misc{liu2025privacy,
  author = {Bangya Liu and Chengpo Yan and Chenghao Jiang and Suman Banerjee and Akarsh Prabhakara},
  title = {{Privacy-Aware Sharing of Raw Spatial Sensor Data for Cooperative Perception}},
  year = {2025},
  eprint = {2512.16265},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.16265},
}
Figure