Awesome

Towards Robustness Certification Against Universal Perturbations

This repository is the official implementation of the ICLR'23 paper "Towards Robustness Certification Against Universal Perturbations". Our goal is to provide the first practical attempt for researchers and practitioners to evaluate the robustness of their models against universal perturbations, especially to universal adversarial perturbations (UAPs) and $l_{\infty}$-norm-bounded backdoors.

Overview

The code in this repository utilizes linear bounds calculated by auto_LiRPA and further computes the certified UP robustness on a batch of data. The calculation of certified robustness can help provide robustness guarantees, identify potential weaknesses in the models and inform steps to improve their robustness.

Requirements

auto_LiRPA (Tested with the February 14, 2023 version)
Gurobipy >= 9.5.1

Usage

Download the example model weights and extract the ./model_weights into the same folder as the code.
Run Jupyter Notebooks for the demos, or load min_correct_with_eps from certi_util.py to calculate the certified UP robustness for your own model and data.

Conclusion

We hope that this repository will serve as a valuable resource for the robustness certification community. By providing a tool to calculate the certified UP robustness, we aim to promote the development of more secure and robust machine learning models.