Adversarial examples for extreme multilabel text classification

Abstract Extreme Multilabel Text Classification (XMTC) is a text classification problem in which, (i) the output space extremely large, (ii) each data point may have multiple positive labels, and (iii) follows strongly imbalanced distribution. With applications recommendation systems automatic tagging of web-scale documents, research on XMTC has been focused improving prediction accuracy dealing with data. However, robustness deep learning based models against adversarial examples largely underexplored. In this paper, we investigate behaviour under attacks. To end, first, define attacks multilabel problems. We categorize attacking classifiers as (a) positive-to-negative, where target label should fall out top-k predicted (b) negative-to-positive, negative be among labels. Then, by experiments APLC-XLNet AttentionXML, show that are highly vulnerable to positive-to-negative but more robust negative-to-positive ones. Furthermore, our success rate an More precisely, tail classes for which attacker can generate samples high similarity actual data-points. overcome problem, explore effect rebalanced loss functions not only do they increase classes, also improve these The code available at https://github.com/xmc-aalto/adv-xmtc .