Deep Partition Aggregation: Robust Defenses Against Poisoning Attacks

Created using ChatSlide

The document explores DPA and SS-DPA methods for robust AI, focusing on MNIST and CIFAR-10 datasets. It delves into adversarial threats like label flipping, details core methodologies including implementation and guarantees, and provides empirical results on certified accuracy and robustness trade-offs. A comparison with historical poisoning defenses contextualizes these approaches within the broader field of ensemble techniques.