基于三維人臉數(shù)據(jù)增強的深度偽造檢測方法
網(wǎng)絡(luò)安全與數(shù)據(jù)治理 9期
王昊冉,,楊敏敏,,王澤源,白亮,,于天元,郭延明
(1.國防科技大學(xué)系統(tǒng)工程學(xué)院,,湖南長沙410073; 2.佳木斯大學(xué)信息電子技術(shù)學(xué)院,,黑龍江佳木斯156100)
摘要: 隨著深度偽造技術(shù)的發(fā)展,深度偽造視頻的制作及傳播變得越來越容易,,給社會帶來了巨大的安全風(fēng)險,,深度偽造檢測算法成為當(dāng)前網(wǎng)絡(luò)安全領(lǐng)域的重要方向。聚焦于提出一種泛化性能更好,、效率更高,、可解釋性更強的深度偽造檢測算法,主要針對DFDC,、FaceForensic++及CelebDF三個深度偽造視頻數(shù)據(jù)集進(jìn)行實驗并以這三個數(shù)據(jù)集集中訓(xùn)練出檢測模型,,首先使用人臉檢測算法MTCNN提取人臉圖像,進(jìn)而將EfficientNet網(wǎng)絡(luò)與Transformer架構(gòu)相結(jié)合作為檢測模型,,通過采用三維人臉數(shù)據(jù)增強,、注意力機(jī)制以及全局局部融合方法對模型進(jìn)行訓(xùn)練和測試。模型在未使用型集成,、知識蒸餾等方法的基礎(chǔ)上,,達(dá)到了與最優(yōu)檢測效果相當(dāng)?shù)臋z測水平。
中圖分類號:KN 90
文獻(xiàn)標(biāo)識碼:A
DOI:10.19358/j.issn.2097-1788.2023.09.003
引用格式:王昊冉,,楊敏敏,,王澤源,等.基于三維人臉數(shù)據(jù)增強的深度偽造檢測方法[J].網(wǎng)絡(luò)安全與數(shù)據(jù)治理,,2023,,42(9):11-20.
文獻(xiàn)標(biāo)識碼:A
DOI:10.19358/j.issn.2097-1788.2023.09.003
引用格式:王昊冉,,楊敏敏,,王澤源,等.基于三維人臉數(shù)據(jù)增強的深度偽造檢測方法[J].網(wǎng)絡(luò)安全與數(shù)據(jù)治理,,2023,,42(9):11-20.
Deepfake detection based on 3D face data augmentation
Wang Haoran1, Yang Minmin2,Wang Zeyuan1,,Bai Liang1,,Yu Tianyuan1,Guo Yanming1
(1.College of System Engineering, National University of Defence Technology, Changsha 410073, China; 2.School of Information and Electonics Technology,, Jiamusi University,Jiamusi 156100, China)
Abstract: With the development of deepfake technology, the production and dissemination of deepfake videos have become increasingly easy, posing significant security risks to society. Therefore, researching deepfake detection algorithms has become an important direction in the field of network security. This paper focuses on proposing a deepfake detection algorithm with better generalization performance, higher efficiency, and stronger interpretability. Experiments are conducted on three deepfake video datasets: DFDC, FaceForensic++, and CelebDF. Firstly, the Multitask Cascaded Convolutional Networks (MTCNN) face detection algorithm is used to extract facial images. Then, the EfficientNet network is combined with the Transformer architecture as the detection model. The model is trained and tested using data augmentation, attention mechanisms, and globallocal fusion methods. Without employing complex model ensembles or knowledge distillation, our model achieves comparable detection performance to stateoftheart methods.
Key words : deep forgery detects;attention-mechanism; data augmentation; neural networks
0 引言
隨著深度學(xué)習(xí)技術(shù)特別是對抗生成網(wǎng)絡(luò)(GAN)技術(shù)的不斷發(fā)展以及互聯(lián)網(wǎng)及個人計算機(jī)的普及,,偽造視頻的濫用也在隨之增長[1]。大量包含虛假政治人物信息的深度偽造視頻在社交媒體上傳播引發(fā)廣泛關(guān)注[2],。準(zhǔn)確鑒別深度偽造視頻,,防止其產(chǎn)生惡劣社會影響成為輿論安全領(lǐng)域一個重要的課題,鑒于此,,國內(nèi)外均采取一定的措施,。2017年,,《新一代人工智能發(fā)展規(guī)劃》經(jīng)國務(wù)院頒布,該規(guī)劃繪制了我國人工智能發(fā)展的宏偉藍(lán)圖[3],。2018年,,美國國會官方定義了“深度偽造”概念,并于當(dāng)月通過了《禁止惡意深度偽造法令》[4],。2019年,,美國國際戰(zhàn)略研究中心(CSIS)針對深度偽造技術(shù)政策發(fā)布簡報。2020年,,美國國防高級研究計劃局(DARPA)為“欺騙逆向工程”項目發(fā)布了一份招標(biāo)文件,,該項目旨在對信息欺騙攻擊的工具鏈進(jìn)行逆向工程。
本文詳細(xì)內(nèi)容請下載:http://forexkbc.com/resource/share/2000005655
作者信息:
王昊冉1,,楊敏敏2,,王澤源1,白亮1,,于天元1,,郭延明1
(1.國防科技大學(xué)系統(tǒng)工程學(xué)院,湖南長沙410073;2.佳木斯大學(xué)信息電子技術(shù)學(xué)院,,黑龍江佳木斯156100)
此內(nèi)容為AET網(wǎng)站原創(chuàng),,未經(jīng)授權(quán)禁止轉(zhuǎn)載。