基于4D-Arnold不等长映射的深度隐写模型参数加密研究

摘要：

隐写模型训练过程中需要大量数据和技术投入，因此隐写模型被窃用将对其所有者造成安全威胁和经济损失。为保护隐写模型，提出了一种基于 4D-Arnold不等长映射的隐写模型参数保护方法。方法采用置乱-扩散策略，首先，置乱阶段通过 4D-Arond 映射对卷积层参数跨卷积核、跨通道置乱。其次，扩散阶段采用相邻参数扩散机制在相邻参数间实现数值扩散并完成参数加密。最后，第三方无法获取任何秘密信息，实现对隐写模型的保护。实验表明，隐写模型加密后提取出的图像在 PSNR，MSE，LPIPS和 SSIM指标以及视觉效果上，显著降低了模型原始性能，模型隐蔽通信功能丧失。此外，所提方法在保证隐写模型加密有效性和安全性的同时，还可以应用于图像分类等其他深度模型的加密保护。

The training process of steganographic models requires a large amount of data and technical investment. When the steganography model is stolen, it will cause security threats and economic losses to its owner. To prevent the theft of deep steganography models, we propose a method for protecting the parameters of the steganography model basing on 4D-Arnold unequal length mapping. The method applies a scrambling-diffusion strategy. Firstly, at the scrambling stage, we scramble the convolutional layer parameters across convolutional cores and channels through 4D-Aronld mapping. Secondly, at the diffusion stage, we design a neighboring parameter diffusion mechanism to achieve numerical difusion between two adiacent parameters and complete the eneryption of the parameters of the deep steganography model, Finally, third parties can't obtain any secret information and we realize the protection of the steganography model. Experiment results show that the method significantly reduces the original performance of the model in terms of objective indicators( PSNR, MSE, LPIPS and SSlM)and visual effects, and the hidden communication funetion of the model is lost. In addition, the proposed method can also be applied to the encryption protection of other deep models such as image classification while ensuring the effectivenes and security of the encryption of the steganography model.

作者：

段新涛，李壮，张恩

Duan Xintao,Li Zhuang,Zhang En

机构地区：

河南师范大学计算机与信息工程学院；河南师范大学教育人工智能与个性化学习河南省重点实验室

引用本文：

段新涛，李壮，张恩。基于4D-Arnold不等长映射的深度隐写模型参数加密研究 [ ] ].河南师范大学学报(自然科学版),2025,53(4):66-73.( Duan Xintao,Li Zhuang,Zhang En, Deep steganography model parameter encryption method based on 4l-Arnold unequal length mapping [ J ].Journal of Henan Normal University( Natural Science Edition),2025,53(4):66-73.D01:10.16366/j.cnki.1000-2367.2024.01.27.0002.)

基金：

国家自然科学基金；河南省高等学校重点科研项目；河南省科技攻关计划

关键词：

AI模型安全；参数加密；4D-Arnold不等长映射；图像隐写模型;卷积神经网络

Al model security; parameter encryption; 4D-Arnold unequal length mapping; image steganography model;convolutional neural network

分类号：

TP309.7

基于4D-Arnold不等长映射的深度隐写模型参数加密研究.pdf