On Single Source Robustness in Deep Fusion Models
Not on the map. Why?
Abstract (arXiv)
Algorithms that fuse multiple input sources benefit from both complementary and shared information. Shared information may provide robustness against faulty or noisy inputs, which is indispensable for safety-critical applications like self-driving cars. We investigate learning fusion algorithms that are robust against noise added to a single source. We first demonstrate that robustness against single source noise is not guaranteed in a linear fusion model. Motivated by this discovery, two possible approaches are proposed to increase robustness: a carefully designed loss with corresponding training algorithms for deep fusion models, and a simple convolutional fusion layer that has a structural advantage in dealing with noise. Experimental results show that both training algorithms and our fusion layer make a deep fusion-based 3D object detector robust against noise applied to a single source, while preserving the original performance on clean data.
Key facts. Cited by 12 papers in the map (data through 7 Oct 2026). First posted on arXiv 11 Jun 2019.
Builds on
No reference list yet.
Built on by
Citing papers per year: 2021: 1; 2022: 3; 2023: 4; 2024: 3; 2025: 0; 2026: 1.
12 papers.
- 2024One Perturbation is Enough: On Generating Universal Adversarial Perturbations against Vision-Language Pre-training Modelscited by 39
- 2022Self-supervised Learning from 100 Million Medical Imagescited by 31
- 2022Defending Multimodal Fusion Models against Single-Source Adversariescited by 20
- 2023Fusion is Not Enough: Single Modal Attacks on Fusion Models for 3D Object Detectioncited by 11
- 2021Sensor Adversarial Traits: Analyzing Robustness of 3D Object Detection Sensor Fusion Modelscited by 5
- 2023What Makes for Robust Multi-Modal Models in the Face of Missing Modalities?cited by 5
- 2024MMCert: Provable Defense against Adversarial Attacks to Multi-modal Modelscited by 4
- 2024Efficient and Effective Universal Adversarial Attack against Vision-Language Pre-training Modelscited by 4
- 2022Training-Free Robust Multimodal Learning via Sample-Wise Jacobian Regularizationcited by 2
- 2023Informative Data Selection with Uncertainty for Multi-modal Object Detectioncited by 2
- 2023Quantifying the robustness of deep multispectral segmentation models against natural perturbations and data poisoningcited by 1
- 2026LAMP: Learning Universal Adversarial Perturbations for Multi-Image Tasks via Pre-trained Modelscited by 0
- 2026LAMP: Learning Universal Adversarial Perturbations for Multi-Image Tasks via Pre-trained Models29 Jan 2026
- 2024Efficient and Effective Universal Adversarial Attack against Vision-Language Pre-training Models15 Oct 2024
- 2024One Perturbation is Enough: On Generating Universal Adversarial Perturbations against Vision-Language Pre-training Models8 Jun 2024
- 2024MMCert: Provable Defense against Adversarial Attacks to Multi-modal Models28 Mar 2024
- 2023What Makes for Robust Multi-Modal Models in the Face of Missing Modalities?10 Oct 2023
- 2023Quantifying the robustness of deep multispectral segmentation models against natural perturbations and data poisoning18 May 2023
- 2023Fusion is Not Enough: Single Modal Attacks on Fusion Models for 3D Object Detection28 Apr 2023
- 2023Informative Data Selection with Uncertainty for Multi-modal Object Detection23 Apr 2023
- 2022Defending Multimodal Fusion Models against Single-Source Adversaries25 Jun 2022
- 2022Training-Free Robust Multimodal Learning via Sample-Wise Jacobian Regularization5 Apr 2022
- 2022Self-supervised Learning from 100 Million Medical Images4 Jan 2022
- 2021Sensor Adversarial Traits: Analyzing Robustness of 3D Object Detection Sensor Fusion Models13 Sep 2021