Adversarial Learning from Crowds

Learning from Crowds (LFC) seeks to induce a high-quality classifier training instances, which are linked range of possible noisy annotations crowdsourcing workers under their various levels skills and own preconditions. Recent studies on LFC focus designing new methods improve the performance trained crowdsourced labeled data. To this day, however, there remain under-explored security aspects systems. In work, we seek bridge gap. We first show that models vulnerable adversarial examples---small changes input data can cause classifiers make prediction mistakes. Second, propose an approach, A-LFC for robust Our empirical results three real-world datasets proposed approach substantially even with existence examples. On average, has 10.05% 11.34% higher test robustness than state-of-the-art in white-box black-box attack settings, respectively.