Underwater image segmentation is crucial for applications such as marine resource exploration and biological research. However, the training process of segmentation models typically requires a large number of images and substantial computational resources. As a result, third-party service resources are often utilized, which poses the risk of image privacy leakage. To address this issue, this paper proposes a privacy-preserving algorithm for underwater images called AFGIPP (Adaptive Fuzzy Gaussian Image Privacy Protection) based on differential privacy. By adding adaptive noise to the luminance channel in the YUV color space, the algorithm effectively protects image information security while retaining the usability of the images. Additionally, to improve the accuracy of underwater image segmentation, we introduce AquaCrossNet, an underwater image segmentation network that innovatively incorporates a Cross-Granularity Complementary Fusion Module (CGCFM). This module enhances the complementarity of multi-granularity features through Transformer. Experiments demonstrate that AFGIPP exhibits superior privacy protection performance. Moreover, AquaCrossNet outperforms eight other methods in multiple objective evaluation metrics and also achieves good segmentation results on images processed by AFGIPP.