载体来源失配下图像隐写分析的前馈控制问题研究

Image steganalysis relies on statistical learning methods which assume that the training and test sets have similar probability distributions. In the real situation, this assumption is too difficult to meet, and the detection accuracy decreased significantly. Therefore, domain adaptation problem, which is also called the cover source mismatch problem in steganalysis, is exposed. The cover source mismatch problem causes instability of steganalytic accuracy, and imposes restriction on the practicality of steganalysis. We've noticed that steganalysis can be viewed as an open-loop control system. Therefore, one important and practical question is that, under the influence of cover source mismatch problem, is it possible to raise the detection accuracy using feedforward control? The image evaluation model used for feedforward control, which helps enrich the existing steganalytic features and make steganalysis more accurate and stable, is the core part of the feedforward control problem. In this paper, we extract novel features which affect the detection rate significantly and improves existing classifier via feedforward control in attempt to determine the presence of steganography. This project aims at the mainstream steganography and steganalysis algorithms. The models we intend to study initially include pixel correlation model, energy model in transformed domain, the model of pixel amounts, etc., and we look forward to finding more models through this project. Our research group has much accumulation for many years in applied research, with the ability and preliminary facilities to conduct this research.

图像隐写分析依赖统计学习手段，需假设训练集与测试集具有相近的概率分布。现实环境中，该假设条件难以满足，准确率明显下降，暴露出领域适应性问题，即载体来源失配问题。这一问题导致隐写分析准确率不稳定，限制了隐写分析的实用。我们注意到，隐写分析可被视为一个开环控制系统。因而，一个重要且实际的问题是：载体来源失配下的图像隐写分析是否存在通过前馈控制提高准确率的可能性。其中，图像评价模型是前馈控制问题的核心部分，有助于丰富已有隐写分析特征，提出更准确更稳定的隐写分析方法。本课题尝试研究图像全局评价模型，在该模型下提取对隐写分析准确率有影响的新特征，通过前馈控制改进现有分类器，判断隐写存在性。课题组针对主流隐写和隐写分析算法，初步确定的模型包括像素相关性模型、变换域能量模型、像素数量模型等，并期待依托本课题发现更多的模型。课题组在所申请研究方向已有多年的积累，具备完成课题的基本能力和研究条件。

图像隐写分析依赖统计学习手段，需假设训练集与测试集具有相近的概率分布。现实环境中，该假设条件难以满足，准确率明显下降，暴露出领域适应性问题，即载体来源失配问题。这一问题导致隐写分析准确率不稳定，限制了隐写分析的实用。本项目研究载体来源失配下的图像隐写分析通过前馈控制提高准确率的相关问题。本课题研究图像全局评价模型，在该模型下提取对隐写分析准确率有影响的新特征，通过前馈控制改进现有分类器，判断隐写存在性。课题组针对主流隐写和隐写分析算法，提出了线性预测误差模型和小波域模型等有效提高空域图像和JPEG图像隐写分析可靠性的方法，还提出了基于假设检验和前馈控制将上述图像模型提取的特征用于提高隐写隐蔽性和隐写分析可靠性的现实方法。相关方法发表在国内外重要期刊会议上，发表论文6篇，获授权发明专利2项，软件著作权6项。研究表明，利用上述方法能有效解决载体来源失配下图像隐写分析的前馈控制问题，从而有助于提高隐写安全性和隐写分析可靠性。