search

Daniel-Liu-c0deb0t / 3D-Neural-Network-Adversarial-Attacks

Research on adversarial attacks and defenses for deep neural network 3D point cloud classifiers like PointNet and PointNet++.
25 stars 8 forks source link

About adversarial training #2

Closed linhaojia13 closed 4 years ago

linhaojia13 commented 4 years ago

Hi @Daniel-Liu-c0deb0t , thank you for your good work! It seems that this repo doesn't implement the adversarial training as depicted in the paper. Do I have a misunderstanding?

Daniel-Liu-c0deb0t commented 4 years ago

This repo is the exact code I used to run experiments for the paper. Adversarial training is done by adding an additional loss with fast gradient L2 to the usual training loss. For PointNet, you can see the adversarial training code here: https://github.com/Daniel-Liu-c0deb0t/3D-Neural-Network-Adversarial-Attacks/blob/master/src/pointnet/train.py#L114. For PointNet++, it is similar.

linhaojia13 commented 4 years ago

It's all my fault that I neglected the modified train.py. Thank you for your prompt reply! @Daniel-Liu-c0deb0t

Daniel-Liu-c0deb0t commented 4 years ago

Hey, if you extend the code and do other experiments, then keep me in the loop. I am interested in what you do.

linhaojia13 commented 4 years ago

Sure, it’s my pleasure.