首页|伪影间共性机理驱动的多域感知社交网络深度伪造视频检测

伪影间共性机理驱动的多域感知社交网络深度伪造视频检测

扫码查看
深度伪造技术在社交网络上的滥用引发了人们对视觉内容真实性与可靠性的严重担忧.已有检测算法未充分考虑社交网络上深度伪造视频的退化现象,导致深度伪造检测性能受以压缩为主的伪影信息干扰与上下文相关信息缺失等挑战性问题的限制.压缩编码与深度伪造生成算法上采样操作会在视频上留下伪影,这些伪影可导致真实视频与深度伪造视频间的细粒度差异.该文通过分析压缩伪影与深度伪造伪影的共性机理,揭示了二者间的结构相似性,为深度伪造检测模型抗压缩鲁棒性的增强提供了可靠理论依据.首先,针对压缩噪声对深度伪造特征的干扰,基于压缩伪影与深度伪造伪影频域表示的结构相似性,设计了频域自适应陷波滤波器以消除特定频带上压缩伪影的干扰.其次,为削弱深度伪造检测模型对未知噪声的敏感,设计了基于残差学习的去噪分支.采用基于注意力机制的特征融合方法增强深度伪造判别特征,结合度量学习策略优化网络模型,实现了具有抗压缩鲁棒性的深度伪造检测.理论分析与实验结果表明,与基线方法相比,该文算法在压缩深度伪造视频上的检测性能具有明显提升,并可作为一种即插即用模型与现有检测方法结合以提高其抗压缩鲁棒性.
Deepfake Video Detection on Social Networks Using Multi-domain Aware Driven by Common Mechanism Analysis Between Artifacts
The misuse of deepfake technology on social networks has raised serious concerns about the authenticity and reliability of visual content.The degradation phenomenon of deepfake videos on social networks has not been adequately considered in existing detection algorithms,resulting in deepfake detection performance being limited by challenging issues such as compression artifacts interference and lack of context-related information.Compression encoding and up-sampling operations in deepfake generation algorithms can leave artifacts on videos,which can result in fine-grained differences between real videos and deepfake videos.The common mechanisms between compression artifacts and deepfake artifacts are analyzed to reveal the structural similarities between them,which provides a reliable theoretical basis for enhancing the robustness of deepfake detection models against compression.Firstly,to address the interference of compression noise on deepfake features,the frequency-domain adaptive notch filter is designed based on the structural similarity of compression artifacts and deepfake artifacts to eliminate the interference of compression artifacts on specific frequency bands.Secondly,the denoising branch based on residual learning is designed to reduce the sensitivity of the deepfake detection model to unknown noise.Additionally,the attention-based feature fusion method is adopted to enhance the discriminative features of deepfakes.Metric learning strategies are adopted to optimize network models,achieving deepfake detection with resistance to compression.Theoretical analysis and experimental results indicate that the detection performance of compressed deepfake videos is significantly enhanced by using the algorithm proposed in this paper.It can be used as a plug-and-play model combined with existing detection methods to enhance their robustness against compression.

Deepfake detectionSocial networkCompressed deepfakeCommon mechanism between artifacts

王艳、孙钦东、荣东柱、汪小雄

展开 >

西安理工大学计算机科学与工程学院 西安 710048

西安交通大学网络空间安全学院 西安 710049

四川数字经济产业发展研究院 成都 610036

深度伪造检测 社交网络 压缩深度伪造 伪影间共性机理

2024

电子与信息学报
中国科学院电子学研究所 国家自然科学基金委员会信息科学部

电子与信息学报

CSTPCD北大核心
影响因子:1.302
ISSN:1009-5896
年,卷(期):2024.46(9)