Disagreement-Regularized Importance Sampling for Adversarial Label Corruption
Researchers have developed a new sub-sampling method called Disagreement-Regularized Importance Sampling (DR-IS) to improve robustness against adversarial label corruption in machine learning. This method leverages the disagreement in loss rankings across independent proxy ensembles to identify and down-weight corrupted data points. DR-IS provides theoretical guarantees on sample concentration and contamination bounds, demonstrating empirical superiority over magnitude-based methods like EL2N, particularly under targeted attacks. AI
IMPACT Enhances machine learning model reliability by providing a robust method to handle noisy or intentionally corrupted labels.