Physical Backdoors for Object Recognition

This simple script uses a custom dataset to train an object recognition model with a physical backdoor.

The 9 classes used in this object recognition dataset are: backpack, coffee mug, cell phone, laptop, purse, running shoe, sunglasses, tennis ball, and water bottle.

The physical backdoor trigger is a smile emoji sticker.

Requirements

tensoflow >= 2.0 cuda >= 10.1

Using the code

1. Clone this repository wherever you choose.

2. Cd into the repository folder and run "pip install -r requirements.txt". This will install all the necessary dependencies.

3. Download the object recognition dataset into the cloned folder. The data can be found at this link.

4. To train a physical backdoored model, simply run the command python3 train.py. You can add flags (specified in train.py) to customize the teacher model, etc.