Security Analysis of Camera-LiDAR Semantic-Level Fusion Against Black-Box Attacks on Autonomous Vehicles

R. Spencer Hallyburton, Yupei Liu, Miroslav Pajic

TL;DR

This paper introduces a novel "frustum attack" that exploits the inherent properties of camera-LiDAR fusion in autonomous vehicles (AVs). The attack successfully compromises a wide range of perception algorithms, including those designed to be robust against simpler LiDAR spoofing, and evades existing defenses, posing a significant threat to AV safety.

What Problem Does This Paper Address?

Autonomous vehicles (AVs) rely heavily on sensor fusion, particularly combining camera and LiDAR data, to perceive their environment and make critical driving decisions. While LiDAR-only perception systems have been shown to be vulnerable to spoofing attacks (where attackers inject fake LiDAR points to mislead the system), the effectiveness of such attacks against more robust camera-LiDAR fusion systems was less understood. This paper addresses the gap in understanding the security of camera-LiDAR fusion against sophisticated, black-box attacks, aiming to demonstrate whether sensor fusion truly enhances security or if new attack vectors can bypass these defenses.

Methods

The core contribution is the introduction of the "frustum attack." This attack leverages the fact that a 2D camera detection defines a 3D "frustum" (a cone-like shape) extending into space, within which the true object's 3D position lies. By injecting spoofed LiDAR points within this frustum, the attacker can create false positives (FPs) or false negatives (FNs) that appear semantically consistent with camera detections, making them much harder to filter out. The attack is "black-box," meaning it doesn't require knowledge of the internal workings of the perception algorithms. The authors also analyze "naive" LiDAR spoofing attacks and evaluate existing defenses against both naive and frustum attacks.

Datasets

  • KITTI Dataset: Used for evaluating perception algorithms and simulating attacks. It provides synchronized camera and LiDAR data with ground truth 2D and 3D bounding boxes.
  • LGSVL Simulator: Used for testing the impact of the frustum attack on an end-to-end AV software stack (Baidu Apollo).

Key Insights and Ideas

  • Sensor Fusion Isn't a Silver Bullet (Against Sophisticated Attacks): While camera-LiDAR fusion offers robustness against naive LiDAR spoofing, it's vulnerable to more context-aware attacks like the frustum attack.
  • The "Frustum" is the Key: By understanding the geometric relationship between camera 2D detections and LiDAR 3D data, attackers can craft highly stealthy and effective attacks.
  • Stealthy by Design: The frustum attack's ability to maintain semantic consistency between camera and LiDAR data makes it bypass existing defenses that rely on detecting inconsistencies.
  • Longitudinal Attacks are Devastating: The attack can be sustained over time, creating adversarial tracks that fool the AV's tracking and control systems, leading to dangerous maneuvers.
  • Minimal Attacker Knowledge Required: The attack is black-box and doesn't need precise knowledge of the target's perception system, making it broadly applicable.

Key Results

  • Frustum Attack Effectiveness: The frustum attack successfully compromises all 8 tested perception algorithms across 3 LiDAR-only and 3 camera-LiDAR fusion architectures.
  • Stealthiness to Defenses: The attack is highly stealthy to existing defenses like CARLO, SVF, and ShadowCatcher, with stealthiness rates often exceeding 80-90%.
  • Robustness to Attack Parameters: The attack is effective even with a relatively small number of spoofed points (tens of points) and can be initiated from longer ranges.
  • Longitudinal Attack Impact: Case studies demonstrate that sustained frustum attacks can lead to dangerous outcomes like unnecessary emergency braking and incorrect speed adjustments in AVs.
  • Naive Attack Resilience: Camera-LiDAR fusion systems showed significant resilience (ASR < 5%) to naive LiDAR spoofing attacks, highlighting the advancement the frustum attack represents.

Limitations and Future Work

  • Dataset Representativeness: The reliance on KITTI, while standard, is acknowledged as a potential limitation due to its relatively small size and may not fully represent real-world driving scenarios. Future work will leverage high-fidelity simulators and more diverse datasets for generalization studies.
  • Optical Engineering and Dynamic Spoofing: While the scenarios for the frustum attack are logistically possible, achieving dynamic tracking and aiming of the spoofer at a moving victim vehicle (required for some scenarios) is an engineering challenge that has not been fully demonstrated. Future work will explore advancements in optical engineering and tracking.
  • Shadow Region Vulnerability: The paper identifies the "shadow region" (behind a target object) as particularly vulnerable. Future work will delve into the low-level behavior of perception DNNs to understand this vulnerability better.
  • Defense Development: The paper highlights the inadequacy of current defenses and calls for the development of new defenses specifically designed to counter the frustum attack.
  • Transferability of Attacks: While the study suggests broad applicability, future work will explicitly investigate the success of transferring specific attack traces between different perception algorithms.

Summary updated

Abstract

To enable safe and reliable decision-making, autonomous vehicles (AVs) feed sensor data to perception algorithms to understand the environment. Sensor fusion with multi-frame tracking is becoming increasingly popular for detecting 3D objects. Thus, in this work, we perform an analysis of camera-LiDAR fusion, in the AV context, under LiDAR spoofing attacks. Recently, LiDAR-only perception was shown vulnerable to LiDAR spoofing attacks; however, we demonstrate these attacks are not capable of disrupting camera-LiDAR fusion. We then define a novel, context-aware attack: frustum attack, and show that out of 8 widely used perception algorithms - across 3 architectures of LiDAR-only and 3 architectures of camera-LiDAR fusion - all are significantly vulnerable to the frustum attack. In addition, we demonstrate that the frustum attack is stealthy to existing defenses against LiDAR spoofing as it preserves consistencies between camera and LiDAR semantics. Finally, we show that the frustum attack can be exercised consistently over time to form stealthy longitudinal attack sequences, compromising the tracking module and creating adverse outcomes on end-to-end AV control.
BibTeX
@inproceedings{hallyburton2022security,
  author = {R. Spencer Hallyburton and Yupei Liu and Miroslav Pajic},
  title = {{Security Analysis of Camera-LiDAR Semantic-Level Fusion Against Black-Box Attacks on Autonomous Vehicles}},
  booktitle = {USENIX Security},
  year = {2022},
  url = {https://openreview.net/forum?id=7DiuJp-ji8B},
}
Figure