search

ebagdasa / backdoors101

Backdoors Framework for Deep Learning and Federated Learning. A light-weight tool to conduct your research on backdoors.
MIT License
337 stars 82 forks source link

How do I perform semantic backdoor attack? #11

Closed michaelTJC96 closed 3 years ago

michaelTJC96 commented 3 years ago

I saw that there are different attacks via the syntehsizer.py, but I cannot find any code related semantic backdoor attack. Is it implemented in the code?

michaelTJC96 commented 3 years ago

Also, I am unable to run the attack against ImageNet because of this problem.

image image

ebagdasa commented 3 years ago

you need to download Imagenet yourself. For the semantic attack on images, I don't have it here yet, but for NLP case as in the blind paper you can refer to this modified transformers repo: https://github.com/ebagdasa/backdoored_transformers

michaelTJC96 commented 3 years ago

from the code, I see that only pattern synthesizer works? What does a pattern synthesizer do? Is it a pixel pattern backdoor as mentioned in your paper, which introduces a series of pixels as a backdoor into the image?

ebagdasa commented 3 years ago

yeah

michaelTJC96 commented 3 years ago

Thank you!