Abstract Community detection algorithms have become essential tools for accurately modeling real-world networks. In particular, the emergence of overlapping community detection techniques has made it possible to identify users’ multiple affiliations, significantly enhancing the analysis of interpersonal relationships. However, this also raises serious privacy concerns, as some users or groups may not wish for their social relationships to be exposed via algorithmic inference. Although several community hiding methods have been proposed to address these issues, existing approaches typically overlook the inherently overlapping characteristics of communities, lack cross-scale adaptability, and exhibit limited interpretability. In this study, we propose a novel overlapping community hiding framework based on constrained graph adversarial training. By integrating trainable layers and a masking mechanism into the adversarial training process, our method effectively achieves multi-scale hiding of overlapping communities, while substantially improving the interpretability of the hiding process. To further enhance the effectiveness of the hiding process, we introduce a novel constraint strategy, termed SAG-NE, into graph adversarial training, which explicitly constrains node representations and symmetric approximate gradients within the same community, thereby increasing node dispersion in both feature and gradient spaces and making it significantly more difficult for existing detection algorithms to uncover the true community structures. Experimental results on multiple real-world and synthetic datasets demonstrate that the proposed framework exhibits robust privacy-preserving performance across different scales.
Yang et al. (2026) studied this question.