Systematically Assessing the Security Risks of AI/ML-enabled Connected Healthcare Systems

Mohammed Elnawawy *, Mohammadreza Hallajiyan *, Gargi Mitra *, Shahrear Iqbal &, and Karthik Pattabiraman *

* University of British Columbia

& National Research Council Canada

This repository is a time-series adaptation of the Universal Robustness Evaluation Toolkit (URET) found at the following link: https://github.com/IBM/URET.git

URET is a tool for generating adversarial samples that can deceive an ML model into mispredicting a particular output.

In this repository, URET is integrated and tested on the Deep Residual Time Series Blood Glucose Forecasting model found at the following link: https://github.com/MLD3/Deep-Residual-Time-Series-Forecasting.git

The victim model was submitted to the Blood Glucose Level Prediction Challenge hosted by Ohio University and evaluated on the OhioT1DM dataset. For more information on the challenge and the dataset, please check the following link: https://sites.google.com/view/kdh-2020/bglp-challenge

Steps to obtain dataset

In the paper titled "Systematically Assessing the Security Risks ofAI/ML-enabled Connected Healthcare Systems", we used the OhioT1DM dataset. The dataset is protected under an NDA and therefore we cannot disclose it publicly. However, it can be obtained from the Ohio University by following the instructions listed here:

https://webpages.charlotte.edu/rbunescu/data/ohiot1dm/OhioT1DM-dataset.html

Steps to set up and run with URET integration

1. Install requirements

python setup.py

2. Activate virtual environment

source myvenv/bin/activate

3. Run the model: python drtf.py <input_directory> <output_directory>

python drtf.py 2018data output_2018

or

python drtf.py 2020data output_2020