Cross-domain restricted learning protects the intellectual property rights of the model by establishing transferability barriers between authorized and unauthorized domains. Existing attack methods for breaking these restrictions require expensive labeled data and model parameter modifications, which disrupts source domain performance. We propose an unsupervised visual translation approach (dubbed as UVT) that bypasses cross-domain restrictions through input-space transformation. Specifically, UVT adopts a UNet-Vision Transformer hybrid generator to visually translate unauthorized domain samples into authorized domain style while preserving semantic content. To further improve performance, we incorporate individual certainty loss and global diversity loss based on the output features of normal samples in the model, ensuring translated samples produce confident yet diverse predictions. Extensive experiments demonstrate that UVT achieves 47.1% performance improvement on unauthorized domains using only 10% of unlabeled data. Meanwhile, the method fully preserves authorized domain accuracy across multiple datasets.