基于特征选择和图卷积表示的JPEG图像隐写者识别

Steganographer identification of JPEG image based on feature selection and graph convolutional representation

针对JPEG图像隐写检测特征维度过高,导致用户间距离计算复杂且隐写者识别性能下降的问题,提出了一种基于特征选择和图卷积表示的JPEG图像隐写者识别方法。首先,提取每个用户图像集的隐写检测特征并度量特征的可分性,从中选取高可分的特征子集;接着,将用户表示为图结构,选取的高可分特征作为图中节点表示,通过训练图卷积神经网络来获得用户特征;最后,考虑类间可分性和类内聚集性,学习到能更大程度捕捉用户差异的表征,提高识别性能。基于常用的BOSSbase-1.01和BOWs图像库的大量实验结果表明,针对利用ns F5、UED、J-UNIWARD等多种主流JPEG隐写方法在图像上嵌入秘密信息的隐写者,所提方法在降低特征维度和计算开销的前提下,多种嵌入比率下的识别准确率均在80.4%以上,并且在低嵌入比率下识别准确率具有明显优势。

Aiming at the problem that the feature dimension of JPEG image steganalysis is too high,which leads to the complexity of distance calculation between users and a decrease in the identification performance of the steganographer,a method for steganographer recognition based on feature selection and graph convolutional representation was proposed.Firstly,the steganalysis features of the user’s images were extracted,and the feature subset with highseparability was selected.Then,the users were represented as a graph,and the features of users were obtained by training the graph convolutional neural network.Finally,because inter-class separability and intra-class aggregation were considered,the features of users that could capture the differences between users were learned.For steganographers who use JPEG steganography,such as nsF5,UED,J-UNIWARD,and so on,to embed secret information in images,the proposed method can reduce the feature dimensions and computing.The identification accuracy of various payloads can reach more than 80.4%,and it has an obvious advantage at the low payload.