In this paper, we assume that we have two types of datasets for classifier design. One is an in-house dataset which is fully available for classifier design as training data. The other is an external dataset which is kept under a very severe privacy preserving policy. We assume that the available information on the external dataset is only the error rate of a presented classifier. No other information is available such as the number of patterns, attribute values of each pattern, and its class label. Thus, the external dataset can be viewed as a black-box model where the error rate is calculated as an output for an input classifier. In this paper, we discuss how such a black-box type dataset can be utilized in fuzzy genetics-based machine leaning (GBML). We use a hybrid fuzzy GBML algorithm where its Michigan-style part is applied to each individual of a Pittsburgh-style part. Since a fuzzy rule-based classifier is an individual in the Pittsburgh-style part, a black-box type dataset can be utilized for fitness evaluation. Through computational experiments, we examine the effect of using a black-box type dataset in comparison with fuzzy rule-based classifiers design only from a fully available dataset.