《计算机应用》唯一官方网站 ›› 2026, Vol. 46 ›› Issue (8): 2515-2523.DOI: 10.11772/j.issn.1001-9081.2025070826

• 网络空间安全 • 上一篇    下一篇

多模态物理先验特征融合的深度伪造检验方法

吕仁堃1, 孙鹏1,2(), 郎宇博1, 郭弘2, 沈喆3, 田迪4   

  1. 1.中国刑事警察学院 公安信息技术与情报学院,沈阳 110854
    2.司法部 司法鉴定重点实验室,上海 200063
    3.沈阳航空航天大学 民航学院,沈阳 110136
    4.辽宁省公安厅 网络安全总队,沈阳 110032
  • 收稿日期:2025-07-23 修回日期:2025-09-25 接受日期:2025-09-26 发布日期:2025-11-05 出版日期:2026-08-10
  • 通讯作者: 孙鹏
  • 作者简介:吕仁堃(2000—),男,山东烟台人,硕士研究生,CCF会员,主要研究方向:深度伪造检验、图像处理
    孙鹏(1978—),男,辽宁沈阳人,教授,博士生导师,博士,主要研究方向:视频侦查、图像处理、多媒体取证
    郎宇博(1989—),男,辽宁沈阳人,讲师,博士研究生,主要研究方向:图像处理、视频/图像伪造检验
    郭弘(1978—),女,上海人,硕士,主要研究方向:电子数据取证、信息安全
    沈喆(1981—),女,辽宁沈阳人,讲师,博士,主要研究方向:图像处理
    田迪(1975—),男,辽宁沈阳人,主要研究方向:网络信息安全。
  • 基金资助:
    国家自然科学基金资助项目(61307016);司法部司法鉴定重点实验室开放课题(KF202317);辽宁省研究生教育教学改革研究项目(LNYJG2023317);沈阳市科技计划项目社会治理科技专项(24-213-3-41);校级重点攻关计划项目(ZDGGJH2507)

Deepfake detection method based on fusion of multi-modal physical prior features

Renkun LYU1, Peng SUN1,2(), Yubo LANG1, Hong GUO2, Zhe SHEN3, Di TIAN4   

  1. 1.College of Public Security Information Technology and Intelligence,Criminal Investigation Police University of China,Shenyang Liaoning 110854,China
    2.Key Laboratory of Forensic Science,Ministry of Justice,Shanghai 200063,China
    3.College of Civil Aviation,Shenyang Aerospace University,Shenyang Liaoning 110136,China
    4.Network Security Corps,Liaoning Provincial Public Security Bureau,Shenyang Liaoning 110032,China
  • Received:2025-07-23 Revised:2025-09-25 Accepted:2025-09-26 Online:2025-11-05 Published:2026-08-10
  • Contact: Peng SUN
  • About author:LYU Renkun, born in 2000, M. S. candidate. His research interests include deepfake detection, image processing.
    LANG Yubo, born in 1989, Ph. D. candidate, lecturer. His research interests include image processing, video/image forgery detection.
    GUO Hong, born in 1978, M. S. Her research interests include digital forensics, information security.
    SHEN Zhe, born in 1981, Ph. D., lecturer. Her research interests include image processing.
    TIAN Di, born in 1975. His research interests include network and information security.
  • Supported by:
    National Natural Science Foundation of China(61307016);Open Project of Key Laboratory of Forensic Science, Ministry of Justice(KF202317);Postgraduate Education and Teaching Reform Research Project of Liaoning Province(LNYJG2023317);Science and Technology Program of Shenyang — Special Project for Social Governance(24-213-3-41);University-Level Key Research Project(ZDGGJH2507)

摘要:

现有深度伪造检验方法多数依据图像像素级线索建模,较少考虑合成过程对伪造图像的影响,尽管取得较好的检验效果,也很难解释检验过程。因此,提出一种多模态物理先验特征融合的深度伪造可解释检验方法。首先,使用光流特征、光照特征、边缘特征和离散余弦变换(DCT)特征分别描述时序视频中的帧间运动差异性、单帧视频中的光照不一致性和边缘伪影信息,得到具有可解释性的多模态物理先验特征;其次,提出多模态混合专家网络,分别构建不同模态的专家子网络,并把子网络在跨模态注意力加权后经门控单元融合,并输入判别网络中实现分类;然后,在判别网络中引入SIAM(Spatial Intersection Attention Module),并将全连接结构替换为KAN(Kolmogorov-Arnold Network)结构;最后,利用多模态物理先验特征分别对不同的专家子网络进行训练,给出对不同输入特征的Shapley值分析,构建事前特征-事后解释的可解释分析框架,为模型推理和预测提供像素级解释。实验结果表明,与CORE(COnsistent REpresentation learning)、SRM(Rich Models for Steganalysis)和UCF(Uncovering Common Features)等算法相比,所提方法在FaceForensics++数据集上的ROC曲线下面积(AUC)和准确率最好,准确率范围为97.35%~98.75%,平均准确率达98.22%,模型可解释性也有较大提升。

关键词: 深度伪造检验, 物理先验特征, 混合专家网络, 可解释性分析

Abstract:

The existing deepfake detection methods mainly model on the basis of pixel-level clues of images, and seldom consider the impact of the synthesis process on the forged images. Although good detection results are achieved, it is difficult to explain the detection process. Therefore, a multi-modal physical prior feature fusion-based explainable detection method for deepfakes was proposed. First, optical flow features, illumination features, edge features and DCT(Discrete Cosine Transform) features were used to describe the inter-frame motion differences in temporal videos, the illumination inconsistency in single-frame videos, and edge artifact information, respectively, so as to obtain multi-modal physical prior features with explainability. Second, a multi-modal mixture of experts network was proposed to construct expert sub-networks for different modalities, and after cross-modal attention weighting, the sub-networks were fused through a gated unit and input into the discriminative network for classification. Third, the SIAM (Spatial Intersection Attention Module) was introduced into the discriminative network, and the fully connected structure was replaced by the KAN (Kolmogorov-Arnold Network) structure. Finally, the multi-modal physical prior features were used to train different expert sub-networks, respectively, and the Shapley value analysis of different input features was given, thereby constructing a pre-feature-post-explanation explainable analysis framework to provide pixel-level explanations for model inference and prediction. Experimental results show that compared with algorithms such as CORE(COnsistent REpresentation learning), SRM(Rich Models for Steganalysis), and UCF(Uncovering Common Features), the proposed method achieves the best performance on AUC and accuracy, with an accuracy range of 97.35% to 98.75% and an average accuracy of 98.22% on FaceForensics++ dataset, and the model’s interpretability also is improved significantly.

Key words: deepfake detection, physical prior feature, mixture of experts network, interpretability analysis

中图分类号: