Generative adversarial networks have demonstrated significant potential in automatically learning image embedding costs. However, the existing U-Net-based generator structure uses the same convolution repeatedly when extracting image features, resulting in a lack of diversity of extracted features and serious information loss with the increase in network depth. To address this issue, this paper introduces a generative adversarial network model based on dual-domain fusion. This model introduces a wavelet convolution branch in the generator to capture various frequency features and spatial information, thereby enriching the amount of information that the model can obtain. Meanwhile, we have also designed a multi-level feature fusion module that efficiently integrates the features obtained from the encoding end, the wavelet convolution branch, and the upsampling decoding end. Through multi-level feature extraction and fusion, the ability of the network to capture image details is greatly improved, and the reconstruction ability of the decoding end is enhanced, optimizing the performance of the generator. We confirmed the effectiveness of the proposed method in resisting steganalyzers through multiple experiments.