网络与信息安全学报 ›› 2022, Vol. 8 ›› Issue (3): 76-86.doi: 10.11959/j.issn.2096-109x.2022028

• 专栏:多媒体内容安全 • 上一篇    下一篇

用于失配隐写分析的对抗子领域自适应网络

章蕾, 王宏霞   

  1. 四川大学网络空间安全学院,四川 成都 610065
  • 修回日期:2022-05-10 出版日期:2022-06-15 发布日期:2022-06-01
  • 作者简介:章蕾(1998− ),女,吉林长春人,四川大学硕士生,主要研究方向为图像隐写分析与深度学习
    王宏霞(1973− ),女,河北赵县人,四川大学教授、博士生导师,主要研究方向为多媒体信息安全、信息隐藏与数字水印、数字取证、智能信息处理
  • 基金资助:
    国家自然科学基金(61972269);四川省科技计划(2022YFG0320);河南省网络空间态势感知重点实验室开放课题(HNTS2022003)

Adversarial subdomain adaptation network for mismatched steganalysis

Lei ZHANG, Hongxia WANG   

  1. School of Cyber Science and Engineering, Sichuan University, Chengdu 610065, China
  • Revised:2022-05-10 Online:2022-06-15 Published:2022-06-01
  • Supported by:
    The National Natural Science Foundation of China(61972269);Sichuan Science and Technology Program(2022YFG0320);Open Foundation of Henan Key Laboratory of Cyberspace Situation Awareness(HNTS2022003)

摘要:

当训练集数据和测试集数据来自不同的载体源时,即在载体源失配的条件下,通常会使一个表现优异的隐写分析器检测准确率下降。在实际应用中,隐写分析人员往往需要处理从互联网上采集的图像。然而,与训练集数据相比,这些可疑图像很可能具有完全不同的捕获和处理历史,导致隐写分析模型可能出现不同程度的检测性能下降,这也是隐写分析工具在现实应用中很难成功部署的原因。为了提高基于深度学习的隐写分析方法的实际应用价值,对测试样本信息加以利用,使用领域自适应方法来解决载体源失配问题,将训练集数据作为源领域,将测试集数据作为目标领域,通过最小化源领域与目标领域之间的特征分布差异来提高隐写分析器在目标领域的检测性能,提出了一种对抗子领域自适应网络(ASAN,adversarial subdomain adaptation network)。一方面从生成特征的角度出发,要求隐写分析模型生成的源领域特征和目标领域特征尽可能相似,使判别器分辨不出特征来自哪一个领域;另一方面从减小域间特征分布差异的角度出发,采用子领域自适应方法来减少相关子领域分布的非期望变化,有效地扩大了载体与载密样本之间的距离,有利于分类精度的提高。通过在多个数据集上对3种隐写算法进行检测,证实了所提方法可以有效地提升模型在数据集失配和算法失配时的检测准确率,减少了失配问题带给模型的负面影响。

关键词: 图像隐写分析, 载体源失配, 对抗学习, 领域自适应

Abstract:

Once data in the training and test sets come from different cover sources, that is, under the condition of cover source mismatch, it usually makes the detection accuracy rate of an outstanding steganalysis model to be reduced.In practical applications, the analyzers need to process images collected from the Internet.However, compared with the training set data, these suspicious images are likely to have completely different capture and processing histories, which may lead to the degradation of steganalysis model.It is also why steganalysis tools are difficult to deploy successfully in the real-world applications.To improve the practical application value of steganalysis methods based on deep learning, test sample information is utilized and domain adaptation method is used to solve the problem of cover source mismatch.Regarding the training set data as the source domain and test set data as the target domain, the detection performance of steganalysis models in the target domain is enhanced by minimizing the discrepancy between the feature distribution of source domain and target domain.ASAN (adversarial subdomain adaptation network) was proposed from the perspective of feature generation on the one hand.The source domain features and target domain features generated by the steganalysis model were required to be as similar as possible, so that the discriminator cannot distinguish which domain the features came from.On the other hand, to reduce the difference of feature distribution between domains, the subdomain adaptation method was adopted to reduce the unexpected change of the distribution of related subdomains.The distance between the cover and stego samples was enlarged effectively to improve the classification accuracy.After testing three steganography algorithms on multiple datasets, it is confirmed that the proposed method can effectively improve the detection accuracy rate of the model in the case of dataset mismatch and algorithm mismatch and it can also reduce the negative impact of the mismatch problem of the model.

Key words: image steganalysis, cover source mismatch, adversarial learning, domain adaptation

中图分类号: 

No Suggested Reading articles found!