Conjunto de Dados para Envenenamento de Modelos Baseados em Self-Training em Fluxo de Dados

In problems with a large volume of unlabeled data, semi-supervisedlearning techniques, such as self-training, are attractive becausethey make full use of the data and do not require extensive labelingof the data, since it is an expensive process. However, using pseudolabelsto train a model indiscriminately can lead to undue changes inthe model’s decision boundary, which can happen unintentionallyor intentionally, such as in malware classification, where attackerswant to classify malicious software as benign. In this paper, wepropose a dataset for poisoning models based on self-training thatsimulates a data stream, intending to evaluate the robustness ofthese models against intentional or unintentional poisoning byunlabeled instances. Our experiments use models from the MOA-SSframework, and show that models that use incremental trainingand prediction confidence as a criterion for using the unlabeledinstance in training are more susceptible to poisoning.