Desensitizing for Improving Corruption Robustness in Point Cloud Classification through Adversarial Training Published in Pattern Recognition (Elsevier), 2025 Reference: JerkyT/DesenAT

Architecture

Backbone: PointNet++ (PointNet2Encoder) with Set Abstraction modules
Framework: OpenPoints/PointNeXt
Defense: Shapley value-based desensitization for adversarial training
Task: 3D point cloud classification (40 classes)
Input: 1024 points × 3 coordinates
Parameters: 1.47M

Results

Metric	Value
ModelNet40 Val OA	90.36%
ModelNet40 Test OA	90.24%
Training Epochs	300
Batch Size	192

Exported Formats

Format	File	Use Case
PyTorch (.pth)	`pytorch/desenat_v1.pth`	Training, fine-tuning
SafeTensors	`pytorch/desenat_v1.safetensors`	Fast loading, safe
ONNX	N/A	Not exportable (custom CUDA FPS/ball_query ops)
TensorRT	Deferred	Generate on target hardware

Custom CUDA Kernels

This module includes 5 optimized CUDA kernels (L4 sm_89):

Chamfer Distance: 230x speedup over PyTorch CPU
Fused Ball Query + Grouping: 318x speedup
EMD Auction Algorithm: 2.4ms for 4×1024 points
GPU Shapley Estimator: 0.38ms Monte Carlo estimation
GPU Free-Form Deformation: 977x speedup over CPU PyGeM

Usage

python

import torch
from safetensors.torch import load_file

# Load weights
state_dict = load_file("pytorch/desenat_v1.safetensors")

# Build model (requires openpoints from DesenAT repo)
from openpoints.models import build_model_from_cfg
from openpoints.utils import EasyConfig
cfg = EasyConfig()
cfg.load("cfgs/modelnet40ply2048/pointnet++.yaml", recursive=True)
cfg.model.criterion_args = {"NAME": "CrossEntropy"}
model = build_model_from_cfg(cfg.model)
model.load_state_dict(state_dict)
model.eval()

# Inference
points = torch.randn(1, 1024, 3).cuda()
data = {"pos": points, "x": points.transpose(1, 2)}
logits = model(data)  # (1, 40)

Training

Hardware: NVIDIA L4 (23GB VRAM)
Framework: PyTorch 2.11 + CUDA 12.8
Optimizer: AdamW (lr=0.001, wd=0.0001)
Scheduler: Warmup + Cosine decay
Config: See configs/baseline_st.toml

Defense Marketplace Value

Blue team counterpart to point cloud attacks. Hardens 3D perception against:

Adversarial point perturbations
Natural sensor corruptions (rain, fog, dust, vibration)
Critical for autonomous vehicles and military ground platforms

License

Apache 2.0 — Robot Flow Labs / AIFLOW LABS LIMITED

ilessio-aiflowlab/DEF-desenat

DEF-desenat — Point Cloud Adversarial Defense via Shapley-Guided Desensitization

Paper