基于GAN的视频隐写算法

Video Steganography Algorithm Based on Generative Adversarial Network

视频隐写术是一种将视频作为嵌入载体实现隐蔽通信的技术。为了解决视频隐写算法中最优修改概率矩阵生成困难的问题,提高信息传输的隐蔽性,提出了一种基于生成对抗网络(generative adversarial network, GAN)的视频隐写算法,该算法包含两种生成对抗网络,分别生成原始载体与修改概率矩阵,前者能够生成视频的静态后景、动态前景与掩模;后者将前景作为隐写生成器的输入,以提高修改概率矩阵的内容自适应性,并利用Tanh-simulator函数拟合最优嵌入函数,促进梯度的反向传播,基于三维卷积网络的隐写分析器作为隐写判别器,它与隐写生成器进行博弈训练以提高载密样本的抗检测性。实验结果表明,视频生成模块不仅能生成逼真的视频,且前景能够代表视频中的时空特征信息,本算法与经典的S-UNIWARD算法相比,在0.05、0.1、0.3 bpc(bit per channel)的嵌入率下,抵抗彩色隐写分析器SCRM检测的能力提高了0.65%~3.26%。

Video steganography is an art of covert communication which uses videos as cover media. In order to solve the problem of having difficulty generating the optimal modified probability matrix in the video steganography, and guarantee remarkable security of information transmission, a generative adversarial network(GAN)-based video steganography algorithm was proposed. There were two generative adversarial networks that generate the original cover and the modified probability map respectively. The first could generate the static background, dynamic foreground and mask. In order to improve the content adaptability of the modified probability map, the foreground was the input of generator in the second, it used the Tanh-simulator function, fit the optimal embedding function to promote the back propagation of the gradient. The steganalyzer based on the 3 D convolutional network was used as its opponent, the discriminator, and the game training is carried out with the generator to improve the performance on resisting steganalysis. The experimental results show that the video generation module can not only generate realistic videos and the generated foreground information, but symbolizes the spatiotemporal feature information. With the embedding rate of 0.05 bpc, 0.1 bpc, 0.3 bpc, the security performance of our proposed algorithm resisting the SCRM steganalysis increases by 0.65%~3.26% than the classic S-UNIWARD algorithm.