Adversarially robust generalization from network crowd noisy labels

Learning from Crowds, also known as LFC, attempts to produce a decent classifier using training instances linked with a range of possibly noisy annotations supplied by crowdsourcing workers with varying personal preconditions. This process often relies heavily on network communications to facilitate the collection and aggregation of these annotations, as data needs to be transmitted between workers, the crowdsourcing platform, and the machine learning system. Recent research on LFC has concentrated on enhancing the performance of classifiers learned using crowdsourced labeled data. Nevertheless, there are still unanswered questions about the adversarial generalization of LFC systems. We further bridge this gap in this work. In an adversarial environment, we formulate the problem of adversarially robust generalization from crowd noisy labels and propose a novel approach, GENIUS, to bolster classifier resilience against adversarial instances. We conduct a comprehensive evaluation of GENIUS on two open and highly adopted benchmark datasets, MGC and Sentiment, demonstrating its superiority over current state-of-the-art methods in the scenarios of white-box and black-box attacks. • Human-labeled data reduce the model classification capability. • Using adversarial data leaves a gap in accuracy between model training and testing. • Solving the bilevel min–max problem improves the model’s accuracy and generalization. • Experimental findings reveal GENIUS exceeds state-of-the-art A-LFC by 21.9%.