Abstract Gradient Training: A Unified Certification Framework for Data Poisoning, Unlearning, and Differential Privacy

The introduction of Abstract Gradient Training (AGT) marks a significant advancement in the field of machine learning, particularly in certifying model robustness against training data perturbations. While the impact of inference-time data perturbations has been well-studied, the certification of models against training data perturbations remains relatively under-explored. AGT addresses this gap by providing a unified framework that encompasses adversarial data poisoning, machine unlearning, and differential privacy. By establishing provable parameter-space bounds, AGT offers a formal methodology for analyzing the behavior of models trained through first-order optimization methods. This framework not only enhances our understanding of model robustness but also sets the stage for future research in ensuring the integrity and reliability of machine learning systems.