Research map / Point clouds / Point cloud classification
CausalPC: Improving the Robustness of Point Cloud Classification by Causal Effect Identification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CausalPC, a model-agnostic defense for point cloud classifiers that applies front-door adjustment with structure extraction (FPS), PU-Net-based reconstruction with random subsampling, and a joint cross-attention head to identify the causal effect and resist diverse adversarial attacks.
Problem
Point cloud classifiers are vulnerable to adversarial perturbations, and existing defenses (adversarial training, input-oriented outlier removal) fail to generalize to diverse and unseen attack types, especially shape-invariant attacks.
Approach
Model point cloud classification with a causal graph whose hidden confounders (adversarial/sensor noise) are handled via front-door adjustment: estimate P(Y|do(X)) by averaging over structural information extracted by FPS and reconstructed/resampled point clouds, classified by a jointly attended pretrained classifier.
Key results
- On ModelNet40 untargeted attacks, DGCNN average accuracy rises from 29.2% (vanilla) to 72.0% with CausalPC, versus 42.4% for the best baseline (DUP-Net).
- On targeted attacks, average attack success rate for PointNet drops from 73.1% to 2.2%, compared with 7.3% for the best baseline DUP-Net.
- Ablation on DGCNN shows structure-only input yields 31.4% average accuracy, reconstructed-point-cloud-only 43.2%, and the combined causal modeling without attention 70.1%, with full model 72.0%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CausalPC to PointNet | PointNet · ModelNet40, untargeted attacks (Avg.) | accuracy: 54.8 -> 67.4 | Helped |
| Add CausalPC to DGCNN | DGCNN · ModelNet40, untargeted attacks (Avg.) | accuracy: 42.4 -> 72.0 | Helped |
| Add CausalPC to PointCNN | PointCNN · ModelNet40, untargeted attacks (Avg.) | accuracy: 64.1 -> 66.6 | Helped |
| Add CausalPC to PointNet | PointNet · ModelNet40, targeted attacks (Avg.) | attack success rate: 7.3 -> 2.2 | Helped |
| Add CausalPC to DGCNN | DGCNN · ModelNet40, targeted attacks (Avg.) | attack success rate: 2.0 -> 0.7 | Helped |
| Add CausalPC to PointCNN | PointCNN · ModelNet40, targeted attacks (Avg.) | attack success rate: 4.5 -> 2.0 | Helped |
| Add the joint attention module (ablation) | CausalPC with structural information and reconstructed point cloud, DGCNN · ModelNet40, untargeted attacks (Avg.) | accuracy: 70.1 -> 72.0 | Helped |
| Use reconstructed point cloud x in addition to structural information z (ablation) | CausalPC without attention, DGCNN · ModelNet40, untargeted attacks (Avg.) | accuracy: 43.2 -> 70.1 | Helped |
| Use structural information z in addition to reconstructed point cloud x (ablation) | CausalPC without attention, DGCNN · ModelNet40, untargeted attacks (Avg.) | accuracy: 31.4 -> 70.1 | Helped |
Surprising findings
- DUP-Net performs worse than vanilla PointCNN on Drop attacks under untargeted settings, since Drop does not create outliers.
- Adversarial training with IFGM examples still yields a near 100% attack success rate against GeoA3 on PointNet and DGCNN.
Limitations
- Approximations in the sub-modules (FPS, PU-Net reconstruction, subsampling) may introduce inaccuracies that create a gap between the causal framework and the implementation.
- No analytical lower bound on robustness is provided.
- Evaluated only on ShapeNet and ModelNet40 with three point-based classifiers and synthetic attacks.
Experimental setup
Datasets: ModelNet40 ShapeNet
Benchmarks: ModelNet40 ShapeNet
Metrics: classification accuracy attack success rate
Baselines: SOR DUP-Net GvG AT PAGN
Builds on: front-door adjustment PU-Net PointNet DGCNN PointCNN
Closest papers by meaning
- CAP: Robust Point Cloud Classification via Semantic and Structural Modeling (CVPR 2023 · similarity 0.89)
- PointCVaR: Risk-Optimized Outlier Removal for Robust 3D Point Cloud Classification (AAAI 2024 · similarity 0.88)
- Robust Structured Declarative Classifiers for 3D Point Clouds: Defending Adversarial Attacks With Implicit Gradients (CVPR 2022 · similarity 0.88)
- Adversarially Robust 3D Point Cloud Recognition Using Self-Supervisions (NeurIPS 2021 · similarity 0.87)
- SMART-PC: Skeletal Model Adaptation for Robust Test-Time Training in Point Clouds (ICML 2025 · similarity 0.85)
- Hide in Thicket: Generating Imperceptible and Rational Adversarial Perturbations on 3D Point Clouds (CVPR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.