面向失配的图像隐写分析研究进展

Progress in cover-source mismatched image steganalysis

尽管隐写分析在实验室环境下取得了显著的进步,但是在实际应用中,由于训练集和测试集的载体来源、隐写算法和嵌入率经常不同,导致隐写分析器性能下降,这种现象称为失配,严重阻碍了隐写分析的实际应用.因此,对目前面向失配问题的主要隐写分析方法进行了分析与总结.根据解决失配问题的思路,将现有失配隐写分析方法分为3类,即设计训练集、取证辅助和无监督领域适应,并对各类方法进行梳理和对比.基于对比结果,探讨了当前基于无监督领域适应的深度隐写分析模型面临的挑战以及未来的发展方向.研究结果表明:基于无监督领域适应的深度隐写分析模型是目前解决失配问题的最有效方案,领域对齐、中间域桥接、对抗学习等是设计该类深度隐写分析模型的主流思想;引入类别等细粒度信息以提高基于无监督领域适应的深度隐写分析模型的性能是未来研究的方向;针对不平衡样本及单/小样本等更恶劣的失配问题的解决方案仍待进一步探索.

Despite the remarkable progress in steganalysis under laboratory settings,the performance of steganalysis systems often declines in practical applications due to differences in the cover source,steganographic algorithms,and embedding rates between training and testing datasets.This phenom-enon,known as“cover-source mismatch”,severely impedes the practical application of steganaly-sis.Therefore,this paper provides an analysis and summary of the main steganalysis methods currently addressing the cover-source mismatch issue.Based on the approach to solving the mismatch issue,existing mismatch steganalysis methods are categorized into three categories:designing training sets,forensics-aided steganalysis,and unsupervised domain adaptation,with each method being reviewed and compared.After comparison,the paper discusses the challenges faced by current deep steganalysis models based on unsupervised domain adaptation and explores future research directions.The research results indicate that deep steganalysis models based on unsupervised domain adaptation are the most effective solution for addressing mismatch issues to date.Domain alignment,intermediate domain bridging,and adversarial learning are the prevailing concepts in designing this type of deep steganalysis model.Introducing fine-grained information,such as class information,to enhance the performance of deep steganalysis models based on unsupervised domain adaptation is a promising di-rection for future research.Further exploration is needed to solve more severe mismatch issues,such as imbalanced samples and single or few shots scenarios.