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

Are there already generated adversarial samples for downloading? #4

Open remindchobits opened 3 years ago

remindchobits commented 3 years ago

Hi! It's a nice work and so do its continuative research! We are now looking into some defence methods but some code are hard to be reproduced with the same performance and the workload is huge. Is there any possibility to download those generated adversarial samples?

Thanks!

Daniel-Liu-c0deb0t commented 3 years ago

Sorry, there aren't. I do not have access to the machine that was used to generate the adv examples, so I cannot upload them. They are massive files so I did not upload them anywhere before. Additionally, this code isn't easy to run directly and it has been a long while since I've touched it.

If you want something cleaner to run and perhaps modify, you should run the code in the continuation research here. It has more detailed instructions on how to run it. I've reimplemented the most important algorithms from this work in the continuation research. It does take a while to run, but you do not need to run every single attack. Some of them are quite slow.

The file you will need to modify is here. You should comment out some of the attacks (for a first try, you can just run iter_l2_attack, it is the fastest one probably). Then, you should modify the test_attacks tuple with the indexes of the attacks to run. In the file, you will see that a couple of defenses will be applied (those defenses are the same as the ones in this work). It should be decently easy for you to add new defenses based on the other defenses I've written. It should be fairly straightforward to implement other attacks from other papers too (right now, the chamfer attack is implemented). More instructions are in the readme.

Let me know if you need anything else.