基于混合深度学习的JEPG图像自适应隐写分析模型和方法研究

Currently, JEPG image steganalysis based on deep learning draw attention in research field. However, the performance of the existing deep learning based steganalysis is far from the researcher’s expectation. Aiming at handling the hard problems when using the deep learning based steganalysis methods to detect the JPEG image adaptive steganography, this proposal conducts research on the steganalysis models, methods and related theories based on hybrid deep learning: (1) We propose an adaptive steganalysis model based on the LSTM and CNN, which can actively remember or adjust the features related to the embedding information at the micro level, to resolve the problem that currently deep learning based steganalysis methods cannot effectively evolve the adaptive stegagalysis features. (2) We propose an effective region selection model based on the attention mechanism and deep reinforcement learning, which can filter the redundant and disadvantageous region for steganlaysis of JPEG images embedded by adaptive steganography with low embedding rate, to handle the problem of the poor accuracy in detecting JPEG images embedded by low embedding rate of adaptive steganography. (3) We propose a method which can search the similar image set based on the filtering features of cover source image learned by deep learning, as well as a method based on fine-tuning the high level layers of neural network in the models proposed in (1) and (2), to handle the problem of the low detection accuracy when the cover source, JPEG image quality, and other factors are mismatched between the test and the training set. The results of the proposal are promising to promote the practical application of deep learning-based steganalysis, and also provides an example for other fields of information security detection.

目前基于深度学习的JPEG图像隐写分析技术得到关注，然而其实际检测效果低于期望。项目针对基于深度学习隐写分析方法在检测JPEG图像自适应隐写中遇到的困难问题，研究基于混合深度学习的隐写分析模型、方法和理论,包括：提出一种基于联合LSTM和CNN的自适应隐写分析模型，在微观层次主动记忆或调整与嵌入信息有关的特征，解决目前基于深度学习的隐写分析不能有效演化自适应隐写分析特征问题；提出一种基于注意力机制和深度强化学习的有效区域选择模型，在宏观层次过滤低嵌入自适应隐写情况下对检测结果冗余和不利的区域，解决低嵌入率下JPEG图像自适应隐写分析困难问题；设计基于深度学习的图像源滤波特征寻找相似图像集方法，以及在高层网络上进行微调方法，解决在图像来源、质量因子等失配情况下JPEG图像自适应隐写分析准确率低的问题。课题成果有助于推进基于深度学习的JPEG图像隐写分析应用，并可运用到其它信息安全检测领域。

目前基于深度学习的JPEG图像隐写分析技术得到关注，然而其实际检测效果仍低于期望，尤其是在低嵌入率或者失配情况下。项目针对基于深度学习的隐写分析方法在检测自适应隐写算法中遇到的若干困难问题开展研究。主要研究内容包括：1. 研究基于混合深度学习的隐写分析模型， 提出了一种基于联合LSTM和CNN的自适应隐写分析模型，在微观层次主动记忆或调整与嵌入信息相关的特征，有效提升了隐写分析的检测准确率。2. 研究低嵌入条件下的JPEG图像自适应隐写分析困难问题 ，提出了一种基于注意力机制和强化学习的有效区域选择的隐写分析模型，在宏观层次过滤对检测结果冗余和不利的区域，有效提升了隐写分析的检测准确率。3. 研究算法失配以及图像源失配问题，提出了一种基于相似度匹配与神经网络微调的方法、基于深层神经网络特征提取的梯度提升树投票集成的方法，提出了NSND特征匹配方法，有效提升了失配情况下的隐写分析的检测准确率。课题成果有助于推进基于深度学习的图像隐写分析应用，并可运用到其他信息安全检测领域。