In this paper, an unsupervised federated quantum generative adversarial network (UF-QGAN) is proposed to handle optimization issues in wireless communications. The proposed UF-QGAN considers unsupervised learning so that it could work without labeling for the training dataset. By adopting the federated framework, the computational load could be distributed between the cloud and edge. As a representative example, the proposed UFQGAN is applied for non-orthogonal multiple access (NOMA). Achievable sum rate and complexity are investigated for the performance measure. Simulation results show that the UF-QGAN, which not requires data labeling, achieved a similar achievable sum rate compared to the supervised learning scheme