基于深度信念网络的图像隐写分析

In this project, we research image steganalysis based on the framework of the feature extraction and pattern recognition. We adopt the theory and method inspired by deep belief network (DBN) and analyze the maximal diversity feature between stego and cover images. We begin this project on supervised learning together with feature selection from the multi-source input, then extend to unsupervised DBN learning of image steganalysis with the low complexity. The main contents are described as follows. 1) In order to improve the recognition rate in the less category, we randomly resample in the imbalance set using Bayesian network theory. 2) Aiming to adaptive image steganography, we build some cost functions with respect to the multi-object optimation. The cascade structure of DBN can approximate the arbitrary cost function and recognize multi-object once. 3) To increase diversity of the DBN output, we choose nonuniform kernel functions in the output layer. The classification performance of the ensemble DBN is better than one of the single DBN. 4) To balance the training complexity and classification rate, we replace the greedy algorithm with generalized recursive least squares over the layer learning and adopt conjugate gradient search as the joint optimization over multi-layer learning. The innovations of this project include that we build the universal unsupervised learning model, and then achieve adaptive DBN training in the imbalance set, moreover, propose the suite of fusion algorithms from the multi-source feature. As the most point, unsupervised DBN learning will be a new way to treat image steganalysis and increase anti-steganography capability.

本项目研究基于图像特征提取和模式分类的隐写分析技术，将深度信念网络(DBN)的理论和方法用于特征分类，通过分析隐写引起的最大差异化特征，结合多源特征筛选，从有监督学习开始逐渐深入，最终实现无监督低复杂度通用隐写分析。研究内容包括：1)利用贝叶斯网理论对非平衡图像集进行随机重采样，提高少数类识别准确率；2)针对基于图像特性的自适应隐写，为多个优化目标依次建立能量函数，用级联结构DBN拟合能量函数，实现多任务识别；3)在输出层选择非一致核函数，增加输出结果的差异性，形成集成DBN，优于单独学习；4)平衡训练复杂度与分类精度的关系，单层学习中用广义递归最小二乘代替贪婪算法；联合优化选择共轭梯度法降低复杂度。项目的创新性在于：建立无监督通用隐写分析模型、实现非平衡集DBN自适应训练、基于DBN实现异构特征的融合。无监督DBN学习将成为隐写分析的一个新方向，对于提高反隐写检测能力将发挥积极作用。

项目对隐写分析的特征精炼和分类融合两项关键技术进行研究，把深度学习中的特征映射、特征表示和特征分类思想运用到现有的隐写分析框架，针对实用隐写分析中存在的特征提取速度慢、含密图像不易获得、训练与测试图像模式不匹配等问题，提出通过定义特征相似度量，利用深度信念网络等深度学习模型思想提高特征的表征能力和分类器的泛化能力。通过空域、变换域联合提取直方图和共生矩阵特征，借助特征筛选、数值优化等理论得到特征的最佳分类结果。在隐写算法未知的情况下，通过后验学习的分层训练和联合优化，结合隐写分析的实际情况，平衡训练模型复杂度与识别准确率的关系，实现通用无监督隐写分析方案，建立可信赖的隐写检测评价机制，为面向实用的隐写分析提供理论支撑。该项目主要工作共发表论文20篇，其中SCI检索论文9篇，EI检索期刊论文9篇，会议论文5篇。