TABOR: A Highly Accurate Approach to Inspecting and Restoring Trojan Backdoors in AI Systems

About

This repository contains partial code implementation of the paper (https://arxiv.org/pdf/1908.01763.pdf). Currently this repo has been written to work on the GTSRB dataset with the 6 Conv + 2 MaxPooling CNN from the original paper.

Dependencies

This codebase is written in tensorflow and tf.keras and has been tested on tensorflow 1.14 and python 3.6.8

Getting Started

1. Clone the TABOR repository

   git clone https://github.com/UsmannK/TABOR.git

2. Download the training data (GTSRB signs)

   cd TABOR
   ./download_data.sh

3. Run the BadNet trainer:

   python3 tabor/train_badnet.py --train --poison-type FF --poison-loc TL --poison-size 8 --epochs 10 --display

   Currently supported options:

   poison type: FF (firefox logo) and whitesquare

   poison location: TL and BR: Top Left and Bottom Right

   poison size: integers

   To train without any poison, exclude the poison-* options

4. Run TABOR on the best model in output/

   python3 tabor/snooper.py --checkpoint output/badnet-FF-10-0.97.hdf5`

   The final mask and pattern will be written to mask.py and pattern.py

Results

TODO! Currently running TABOR for 500 epochs as in the paper takes about 4 hours on my hardware so generating the full results table is still upcoming.

TODOs:

• Implement BadNets
• Implement image generation and basic categorical crossentropy optimization function
• Implement all regularization terms from the TABOR paper
• Add the outlier detection metric from section 4.2.4 from the TABOR paper
• Generate a results table