基于变分贝叶斯框架的抗去同步攻击数字水印系统研究

As the bottleneck of digital watermarking, desynchronization attack and security attack are recognized as the most challenging problem. The well known exhaustive searching, invariance domain and data-aided schemes are not optimal in the sense of universal decoding, which may also leave security holes in the system.The high computational complexity of the joint estimation and decoding coupled with its inability in model selection prevent its practical usage. We base our research of this problem on the framework of variational Bayesian inference. When combined with code-aided joint estimation and decoding, this framework can provide us low complexity universal decoder. The security level of this framework will also be explored. In the first step, we explore the condition of optimality for joint estimation and decoding when the parameter space is continuous under the framework of Bayesian learning. Then under the framework of variational Bayes, we design low complexity iterative estimation and decoding algorithm based on graphical models and junction tree. We will also analyze the theoretical performance of joint estimation and decoding. The new algorithm will be tested and compared to the invariance domain method and the method based on EM algorithm. Finally, we will explore the theoretical security of this framework and construct practical attacking algorithms, based on which we can improve our algorithm. Hopefully we may obtain low complexity realization of the optimal universal watermark decoder for specific desynchronization models and disclose its performance in terms of estimation, decoding and security. Our research may promote the application of watermarking in copyright protection and authentication. This research may also benefit the development of communication systems with unknown channel statistic, for example, wireless communication of high speed vehicles.

去同步攻击和针对水印密钥的安全性攻击是数字水印的技术瓶颈，为水印研究领域公认的最具挑战性的难点。常见的穷举、不变域、数据辅助方案都不具有通用解码意义上的最优性，且带来安全问题；联合估计解码计算复杂性高且无法确定攻击模型的阶。本课题提出在贝叶斯学习框架下，基于变分贝叶斯推理，采用编码辅助方法，联合进行去同步攻击参数估计和水印解码，以获得高效的抗去同步攻击水印解码算法，并探明该类结构的安全机制。首先从通用解码角度探索贝叶斯学习框架下联合估计解码的最优性条件；然后以变分贝叶斯为工具，研究图上的低复杂度迭代解码算法，对攻击估计性能和解码性能作理论分析；最后，探求上述框架的理论安全性，构建攻击算法，据此改进安全性能。通过本研究，可望获得针对特定攻击模型的低复杂度通用水印解码器，在理论上揭示其攻击估计、解码和安全性能。将促进数字水印在版权保护、身份认证等领域的应用，推动未知信道特性下通信领域的发展。

在针对水印系统的攻击中，去同步攻击具有攻击效果强、引起的感知失真小以及对抗困难的特点，是近年来水印技术发展的绊脚石。本项研究从数字通信信道建模的角度研究这种具有未知参数攻击信道下的最优嵌入、信道估计和解码问题。主要考虑了幅度缩放和滤波这两种攻击信道模型。针对该问题我们研究了如下内容：1）针对幅度缩放参数估计和水印符号解码精度，研究提高精度的方法；2）降低上述高精度算法的计算复杂性的方法，探索建立因子图上的EM算法等变分优化算法；3）在最优Wiener滤波攻击情况下的水印嵌入策略。通过一年的研究，获得了如下结果：1）针对研究内容一，获得了以LDPC码辅助和因子图上消息传递为基础的编码辅助参数估计和解码算法，该算法的比特错误率比Moulin等人基于重复码的算法更接近脏纸代码香农限；2）在变分优化的框架下，获得了以因子图EM算法为基础的、相关分块衰落信道下的低复杂度估计解码算法；3）获得了在Wiener最优滤波攻击情况下，针对非宽平稳载体模型的最优谱条件。上述成果表明图模型上的变分贝叶斯方法在抗去同步攻击中具有优势，为下一步探究更复杂攻击模型下的变分贝叶斯算法提供了垫脚石。