COMMIT: Certifying Robustness of Multi-Sensor Fusion Systems Against Semantic Attacks

Authors: Zijian Huang, Wenda Chu, Linyi Li, Chejian Xu, Bo Li

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we evaluate the efficacy of COMMIT in different settings and provide a comprehensive benchmark of certified robustness for different MSF models using the CARLA simulation platform.
Researcher Affiliation Academia 1University of Michigan 2California Institute of Technology 3Simon Fraser University 4University of Illinois Urbana-Champaign EMAIL, EMAIL, linyi EMAIL, EMAIL, EMAIL
Pseudocode No No explicit pseudocode or algorithm blocks are provided in the main text.
Open Source Code No The benchmark will be open source upon acceptance and will be continuously expanding to evaluate more AD systems.
Open Datasets Yes We construct a diverse dataset leveraging the CARLA simulator (Dosovitskiy et al. 2017). We generate our KITTI-CARLA dataset (Deschaud 2021) with 5, 000 frames in CARLA Town01 with 50 pedestrians and 100 vehicles randomly spawned, in which 3, 500 frames are used for training and 1, 500 frames are used for testing.
Dataset Splits Yes We generate our KITTI-CARLA dataset (Deschaud 2021) with 5, 000 frames in CARLA Town01 with 50 pedestrians and 100 vehicles randomly spawned, in which 3, 500 frames are used for training and 1, 500 frames are used for testing.
Hardware Specification No No specific hardware details (e.g., exact GPU/CPU models, processor types, or memory amounts) used for running experiments are provided in the main text.
Software Dependencies No No specific ancillary software details with version numbers (e.g., library or solver names like Python 3.8, PyTorch 1.9) are provided in the main text.
Experiment Setup Yes To make the models adapt with Gaussian noise smoothed data, we train two sets of models with Gaussian augmentation (Cohen, Rosenfeld, and Kolter 2019) using noise variance ς = 0.25 and ς = 0.5. For the ease of robustness certification, for rotation certification, we use models trained with ς = 0.25 to construct smoothed models; for shifting certification, we use models trained with ς = 0.5. Note that our framework allows using different ς and sample strategies for image and point cloud data. We set the overall confidence of certification to be 95%, aligning with the setting in (Kang et al. 2022).