Mask Matrix-Based Two-Stage Model Function Hiding Algorithm

随着人工智能技术的快速发展，深度神经网络模型已成为重要的数字资产，保护其版权和隐蔽传输成为关键问题。传统的模型水印技术和主动保护方案虽然在一定程度上能够防止模型被盗用，但仍然存在隐蔽性差、性能下降等问题。为此，本文提出了一种基于掩码矩阵的两阶段模型功能隐藏算法，旨在解决深度神经网络模型在公共信道中的隐蔽传输问题。该方法通过两阶段的设计，能够在解码阶段对模型做到无损恢复。算法通过生成掩码矩阵隐藏秘密任务，同时引入参数统计损失约束，最小化掩码矩阵对模型参数分布的影响，提高传输过程中的隐蔽性。实验结果表明，提出方法在不同结构模型上都有优秀的表现，模型加密前后的KL散度平均值为0.0044，秘密任务恢复后的平均准确率可以达到93.08%。所提算法为DNN模型的安全隐蔽传输提供了有效的解决方案，具有广泛的应用前景。With the rapid development of artificial intelligence technology, deep neural network models have become important digital assets, making the protection of their copyright and secure transmission critical issues. Although traditional model watermarking techniques and active protection schemes can prevent model theft, they still face problems such as poor concealment and performance degradation. To address this, this paper proposes a two-stage model functionality hiding algorithm based on a mask matrix, aiming to solve the problem of secure transmission of deep neural network models over public channels. The proposed method enables lossless recovery of the model during the decoding stage through a two-stage design. The algorithm generates a mask matrix to hide the secret task while introducing a parameter statistical loss constraint to minimize the impact of the mask matrix on the model’s parameter distribution, thus improving the concealment during transmission. Experimental results show that the proposed method performs excellently across models with different architectures, with an average KL divergence of 0.0044 before and after model encryption, and the average accuracy of the recovered secret task reaches 93.08%. The proposed algorithm provides an effective solution for the secure and concealed transmission of DNN models, with broad application prospects.