Keeping up with dynamic attackers: Certifying robustness to adaptive online data poisoning

=====================================================

Description

---

This project aims to certify the robustness of machine learning models against adaptive online data poisoning attacks.

Installation

---

To install the dependencies, run the following command: pip install -r requirements.txt

General Certificate Abstract Class

The generalized_certificate.py module provides an abstract class for computing certificates and using them to meta-learn a robust learning algorithm. This class serves as a foundation for implementing specific certification methods.

Example Implementations

MeanEstimation/mean_estimation_toy.py: Provides a concrete implementation of the certificate abstract class for the mean estimation task. Classification/classification_toy.py: Provides a concrete implementation of the certificate abstract class for the classification task.

Mean Estimation

To perform mean estimation, follow these steps:

1. Navigate to the MeanEstimation directory: cd MeanEstimation
2. Run the following script: python3 mean_estimation.py

Classification

To perform classification, follow these steps:

1. Download the MNIST dataset (or any corresponding dataset) and store the dataset in the data folder.
2. Navigate to the Classification directory: cd Classification
3. Run the following script: python3 mnist.py