NEEDLE

This is the repository for paper NEEDLE: Towards Non-invertible Backdoor Attack to Deep Learning Models.
See the paper in arxiv (Coming soon...).

Dependences

The code is run on: Python3.6, tensorflow=1.13.1, keras=2.2.4, numpy, scikit-learn, opencv-python.

How to use this repository

We provide codes on CIFAR-10 dataset.
We provide several partitioning methods:

• K-means
  
• (Coming soon...) GMM
  
• (Coming soon...) Color classifier
  
• (Coming soon...) Sub-labeling
  

And trigger patterns:

• Color patches
  
• (Coming soon...) Logos
  

Prepare

Create checkpoint folder: ./checkpoints.
Download color classifier from Keras-Color-Classifier into ./checkpoints.

Train a feature encoder

Train a feature encoder: python TrainEncoder.py.

Perform NEEDLE

Typing in python NEEDLE.py to inject the backdoor.
Parameters about partitioning secrets are listed in the top of this file. Create your own backdoor!
Customize your partitioning algorithm using Trigger focusing in TriggerFocus.py.
Customize your trigger patterns by modifying TriggerPattern.py.