面向深度伪造检测的高效自解释图神经网络

打开文本图片集

中图分类号：TP391.41 文献标志码：A 文章编号：1001-3695（2026）04-032-1229-09

doi：10.19734/j.issn.1001-3695.2025.07.0209

Efficient and self-explainable graph neural networks for deepfake detection

Lyu Renkun¹，Sun Peng1，2†，Lang Yubo’，Shen Zh ：3 ，Meng Hui1， Zhou Chunbing1 （1.Dept.ofPublicrityfoatioTeolndtelc，CrialIesigaticeUniersityfin China;2.KeyLaboforensic，MinistryofJustice，Shangha2o，China；3.CiilAviationClge，ShenyngAerospaceUnieity Shenyang110135，China）

Abstract：Toaddress thelimitations of existingdeeplearning-based methods indeepfakedetection—namelysuboptimalperformanceandlowinterpretability，thispaperproposedamethodbasedonaneicientandself-explainablegrapheuraletwork （SE-EffGNN）.Theapproach consistedof three components：graph structureconstruction，model prediction，andexplanatory analysis.Firstlyitintegratedpriorknowledgeandlocalspatialinformationtomodelinputvideosassemanticallgudedgaph structures，enhancingthesensitivitytoforgerytraces.Next，itdesignedSE-EfGNNtoencodenodefeaturesacross multipledimensions，updatenoderepresentations viacombined distance-weighted aggregationand adjacency-edge agregation，and incorporatechannel gatingandnodeatentionmechanisms tocapturediscriminativefeaturesforclasification.Finallyitvisualied parametersfromthedistance weighting，featureencoding，andnodeattention modules toprovideself-explanatorycapability.Experimentalresultsshowthatthe proposed method achieves anaverage AUCofO.9946on mainstream datasets.The visualizationoflearnedparametersofersclearinterpretativesupportconfirming thatSE-EfGNNmaintains highetectionperforance while demonstrating excellent explainability.

Key words：deepfake detection；graph neural network；self-explainable analysis；prior information

0引言

随着人工智能技术的飞速发展，深度伪造技术作为一种基于生成对抗网络1和变分自编码器[2等生成模型的衍生技术，能够根据使用者的需求生成高度逼真的伪造图像和视频。（剩余22072字）