Unrelabel

This tool provides a extensive framework for testing and defending machine learning models against data poisoning attacks and backdoors by allowing users to simulate various poisoning methods, analyze their impacts on model accuracy and behavior, and apply multiple defense strategies such as regularization, data sanitization, ensemble methods, and certified partitioning, all while maintaining user data privacy through local execution. Its main features include automatic attack detection, customizable poison injection techniques like backdoors and label flipping, behavioral integrity assessment through canaries, and an array of defense mechanisms with configurable parameters, thereby solving the problem of model vulnerability to malicious data manipulation for data scientists, ML engineers, security researchers, and organizations aiming to ensure model robustness and integrity in sensitive applications.

Unrelabel can be found in Machine Learning Model Deployment categories.