Extending Adversarial Attacks and Defenses to Deep 3D Point Cloud Classifiers
CSEF · 2019 Mathematical Sciences
Overview
Objectives Testing 3D point cloud neural networks to see if they are susceptible to adversarial attacks like previous results on 2D images, proposing new algorithms for both attacking and defending 3D point clouds, and explaining the effectiveness of the attacks and defenses to better understand the nature of 3D point cloud classifiers. Methods Two well-known networks were evaluated: PointNet and PointNet++. They were trained and tested with the ModelNet-40 and smaller ModelNet-Unique dataset. Previously proposed adversarial attacks and defenses were implemented and tested. Also, new algorithms for attacking networks using methods that make use of intrinsic properties of 3D space and the 3D network architectures are proposed. Results 3D point cloud neural networks were found to be similarly susceptible against adversarial attacks compared to 2D image classifiers, and human-imperceptible perturbations can be generated for 3D point clouds. The defense algorithms that I proposed were also effective. 3D point cloud networks were found to exhibit a new "gradient hiding" phenomenon, which allows the correct shape of objects to be hidden from adversarial attacks that require gradients. Conclusions Though 3D point cloud classifying neural networks are weak against adversarial attacks, gradient hiding and other intrinsic properties of 3D point clouds and the network architectures allow 3D point cloud classifiers to be more easily defensible than 2D image classifiers. This means that 3D point cloud classifiers should be favored over 2D image classifiers in tasks that require visual data, like autonomous driving. Also, the gradient hiding theory leads to a better understanding of how 3D neural networks behave against adversarial perturbations.
Summary statement
Examining the behavior of 3D point cloud classifiers against adversarial attacks and attempting to defend them.
Help received
Ronald Yu and Hao Su (both from Department of CSE, UCSD) provided the topic for me to work on, and GPUs for running experiments. I also conversed with Ronald Yu on the general designs of the project. All code were written by me. Almost all algorithm ideas were designed by me.
Competition history
- CSEF 2019
Resources
Related projects
ISEF · 2021
Limited Query Black-box Adversarial Attacks in the Real World
ISEF · 2019
Protection of Deep Neural Networks against Adversarial Attacks with Application to Facial Recognition
ISEF · 2021
A Biologically Inspired Game Theoretic Adversarial Training Method
ISEF · 2024
AdvMed: Detecting Adversarial Attacks in Medical Deep Learning Systems
CSEF · 2018
Analyzing 2D-Convolutional Neural Network's Ability to Learn Obstacle Avoidance in Self-Driving Applications
JSHS · 2024
Adversarial AI: Poking Holes in Classification Algorithms
ISEF · 2023
Increasing Computer Vision Models Interpretability
ISEF · 2026
Detecting Outliers in 3D Point Clouds Using Higher-Order Geometric Cycles for Accurate 3D Reconstruction
Closest projects by meaning, across every fair and year in the corpus.
Browse more like this
Source: California Science & Engineering Fair public projects