MOBA: A Material-Oriented Backdoor Attack against LiDAR-based 3D Object Detection Systems

Saket S. Chaturvedi, Gaurav Bagwe, Lan Zhang, Pan He, Xiaoyong Yuan · 2025

Read MOBA: A Material-Oriented Backdoor Attack against LiDAR-based 3D Object Detection Systems by Saket S. Chaturvedi, Gaurav Bagwe, Lan Zhang, Pan He, Xiaoyong Yuan (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-09999.txt.

Introduction

Saket S. Chaturvedi, Gaurav Bagwe, Lan Zhang, Pan He, and Xiaoyong Yuan introduce MOBA, a backdoor attack on LiDAR-based 3D object detection that is designed to work in the physical world. Backdoors hide malicious behavior in a model during training, but digital triggers often fail outdoors because they ignore how real materials reflect LiDAR pulses. Physically built triggers, in turn, are often unoptimized.

MOBA models the trigger's material directly, selecting titanium dioxide for its strong diffuse reflectivity and weather resilience. A simulation pipeline approximates a standard reflectance model in an angle-independent way and scales by distance so the digital trigger mimics the physical one. Across leading LiDAR and camera-LiDAR fusion detectors, the attack succeeded at a rate above ninety-three percent, beating prior methods by over forty points.

Catalog record

CREATORSaket S. Chaturvedi, Gaurav Bagwe, Lan Zhang, Pan He, Xiaoyong Yuan
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE38 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-09999

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

MOBA: A Material-Oriented Backdoor Attack against LiDAR-based 3D Object Detection Systems Saket S. Chaturvedi, Gaurav Bagwe, Lan Zhang, Pan He, Xiaoyong Yuan Abstract LiDAR-based 3D object detection is widely used in safety-critical systems. However, these systems remain vulnerable to backdoor attacks that embed hidden malicious behaviors during training. A key limitation of existing backdoor attacks is their lack of physical realizability, primarily due to the digital-to-physical domain gap. Digital triggers often fail in real-world settings because they overlook material-dependent LiDAR reflection properties. On the other hand, physically constructed triggers are often unoptimized, leading to low effectiveness or easy detectability. This paper introduces Material-Oriented Backdoor Attack (MOBA), a novel framework that bridges the digital–physical gap by explicitly modeling the material properties of real-world triggers. MOBA tackles two key challenges in physical backdoor design: 1) robustness of the trigger material under diverse environmental conditions, 2) alignment between the physical trigger’s behavior and its digital simulation. First, we propose a systematic approach to selecting robust trigger materials, identifying titanium dioxide (TiO 2 ) for its high diffuse reflectivity and environmental resilience. Second, to ensure the digital trigger accurately mimics the physical behavior of the material-based trigger, we develop a novel simulation pipeline that features: (1) an angle-independent approximation of the Oren–Nayar BRDF model to generate realistic LiDAR intensities, and (2) a distance-aware scaling mechanism to maintain spatial consistency across varying depths. We conduct extensive experiments on state-of-the-art LiDAR-based and Camera-LiDAR fusion models, showing that MOBA achieves a 93.50% attack success rate, outperforming prior methods by over 41%. Our work reveals a new class of physically realizable threats and underscores the urgent need for defenses that account for material-level properties in real-world environments. 1 Introduction LiDAR-based 3D object detection has emerged as a cornerstone of modern autonomous driving systems, enabling vehicles to accurately perceive and localize surrounding objects such as cars, pedestrians, and cyclists ( 19 ) . Powered by deep neural networks, these systems achieve remarkable performance but also inherit vulnerabilities from data-driven learning pipelines–most notably, backdoor attacks ( 6 ) . In backdoor attacks, an adversary poisons a small subset of training data with a hidden trigger pattern, causing the trained model to exhibit targeted misbehavior ( e.g . , false detections or missed objects) when the trigger appears ( 3 ; 29 ; 10 ) . Recent work has demonstrated that LiDAR-based 3D object detectors are highly vulnerable to such attacks ( 29 ; 10 ) . Consider a representative threat scenario: an adversary injects examples of vehicles containing a specific LiDAR-visible trigger ( e.g . , a large ball placed on the roof of a vehicle ( 29 ) ). Once trained on the poisoned data, the deployed model behaves normally under benign conditions. However, when an attacker places the physical trigger on their own vehicle, nearby vehicles equipped with the compromised model fail to detect the vehicle accurately, leading to potentially catastrophic consequences such as intentional collisions or traffic manipulation. This threat surface has further expanded to include multi-modal systems, such as Camera-LiDAR fusion models ( 23 ; 4 ) , further increasing the real-world risk. Figure 1: Overview of MOBA . MOBA injects a digitally simulated, BRDF-informed trigger into a small subset of training data to poison the model. MOBA leverages material reflectance modeling to ensure that the trigger is physically realizable and effective under real-world conditions. When the physical trigger is deployed at inference time, the backdoored model produces adversarial predictions ( e.g . , shrunk or missing bounding boxes). Despite growing attention, a critical technical gap remains: existing digital backdoor attacks are rarely physically realizable, while existing physical attacks are either hand-crafted or poorly optimized, underestimating the threats of backdoor attacks. For instance, prior physical attacks rely on manually placed object-like point clusters without optimization, either 1) producing sparse LiDAR returns that are unreliable across varying conditions ( e.g . , diverse angles and distances) ( 10 ) or 2) deploying with large-size triggers to ensure sufficient LiDAR returns, which is easy to detect ( 29 ) . Conversely, digitally optimized triggers assume ideal sensor responses but ignore the physical characteristics of real-world LiDAR sensing, such as material reflectivity, and surface angle, which leads to significant performance degradation when deployed. To address the gap between digital backdoor optimization and physical realizability, we propose a Material-Oriented Backdoor Attack (MOBA) that introduces a physics-driven methodology for constructing robust and generalizable LiDAR triggers. The core novelty of MOBA lies in its integration of material reflectance modeling, grounded in optical physics, into both the selection and digital simulation of backdoor triggers. Unlike prior works that use hand-crafted or heuristically optimized digital triggers, MOBA formulates a quantitative material selection objective that balances specular and diffuse reflectance under real-world LiDAR wavelengths. We identify titanium dioxide as an optimal off-the-shelf trigger material due to its strong diffuse reflectivity, resilience to environmental variation ( e.g . , dust, rain, wear), low cost, and ease of application via standard coatings. Moreover, to simulate its behavior accurately in training, MOBA introduces a physically grounded, angle- and distance-robust simulation strategy: we derive an angle-independent approximation of the Oren–Nayar BRDF ( 14 ) , a widely used reflectance model for rough, diffuse surfaces, to model the reflectance across unknown incident angles, and design a distance-aware scaling mechanism to maintain the trigger’s spatial consistency in LiDAR point clouds at varying depths. Together, these designs allow MOBA to produce physically plausible triggers that are effective, stealthy, and robust across diverse real-world deployment scenarios. To the best of our knowledge, this is the first work to develop a material-oriented physical backdoor attack targeting state-of-the-art LiDAR-based 3D object detection systems. We conduct extensive evaluations using real-world data collected in outdoor environments, demonstrating that our attack consistently achieves high success rates under diverse viewing conditions. To further evaluate generalizability, we extend our analysis from LiDAR-only models to Camera-LiDAR fusion models, showing that the attack remains effective in multi-modal perception systems. Our contributions are summarized as follows: • We propose MOBA, the first material-oriented backdoor attack that models physically grounded LiDAR reflectivity to generate physically realizable yet highly effective backdoor triggers. • We introduce a robust material selection strategy and identify titanium dioxide as an ideal candidate for its high reflectivity and consistent LiDAR response under varying conditions. • To overcome the challenge of unknown incident angles in LiDAR scans, we derive a novel approximation of the Oren–Nayar model that marginalizes over angular dependencies, allowing intensity values to be simulated without geometric priors. • We propose a depth-robust design that ensures the spatial consistency of the digital trigger under varying distances, mitigating domain shift caused by LiDAR sparsity. • We conduct extensive experiments using collected physical LiDAR-Camera data and show that MOBA achieves a high Attack Success Rate (ASR) of 93.50%, outperforming state-of-the-art physical backdoor attacks by over 41% on average. 2 Related Works 2.1 LiDAR-based 3D Object detection LiDAR sensors are integral to autonomous driving, providing precise depth and spatial measurements that are critical for robust 3D object detection. Architectures for processing LiDAR data are diverse, generally falling into three categories. Voxel-based methods like VoxelNet ( 30 ) discretize the point cloud into a structured voxel grid, enabling the use of efficient 3D

Read the complete text