Outlier detection is an important topic in machine learning and has been usedin a wide range of applications. In this paper, we approach outlier detectionas a binary-classification issue by sampling potential outliers from a uniformreference distribution. However, due to the sparsity of data inhigh-dimensional space, a limited number of potential outliers may fail toprovide sufficient information to assist the classifier in describing aboundary that can separate outliers from normal data effectively. To addressthis, we propose a novel Single-Objective Generative Adversarial ActiveLearning (SO-GAAL) method for outlier detection, which can directly generateinformative potential outliers based on the mini-max game between a generatorand a discriminator. Moreover, to prevent the generator from falling into themode collapsing problem, the stop node of training should be determined whenSO-GAAL is able to provide sufficient information. But without any priorinformation, it is extremely difficult for SO-GAAL. Therefore, we expand thenetwork structure of SO-GAAL from a single generator to multiple generatorswith different objectives (MO-GAAL), which can generate a reasonable referencedistribution for the whole dataset. We empirically compare the proposedapproach with several state-of-the-art outlier detection methods on bothsynthetic and real-world datasets. The results show that MO-GAAL outperformsits competitors in the majority of cases, especially for datasets with variouscluster types or high irrelevant variable ratio.