基于卷積LSTM的視頻中Deepfake檢測(cè)方法
信息技術(shù)與網(wǎng)絡(luò)安全
李永強(qiáng),,白 天
(中國(guó)科學(xué)技術(shù)大學(xué) 軟件學(xué)院,,安徽 合肥230026)
摘要: 以Deepfake為代表的偽造人臉技術(shù),使用少量的人臉數(shù)據(jù)就能將視頻中的人臉替換成為目標(biāo)人臉,,從而達(dá)到偽造視頻的目的,。此類技術(shù)的濫用將帶來惡劣的社會(huì)影響,,需要使用檢測(cè)技術(shù)加以制裁。針對(duì)這一問題,,已有若干檢測(cè)算法被提出?,F(xiàn)有方法具有一定局限性,單幀檢測(cè)算法忽略了Deepfake動(dòng)態(tài)缺陷,;當(dāng)數(shù)據(jù)存在缺陷時(shí),,模型可能會(huì)陷入“學(xué)會(huì)特定臉”的陷阱中。提出了一種對(duì)視頻數(shù)據(jù)中的Deepfake檢測(cè)方法,,使用結(jié)合CNN和LSTM的卷積LSTM,,判斷視頻真?zhèn)巍L岢隽艘环N基于人臉特征點(diǎn)的cutout方法,能抑制網(wǎng)絡(luò)學(xué)會(huì)特定臉,。實(shí)驗(yàn)表明,,在不同場(chǎng)景下,準(zhǔn)確度對(duì)比基準(zhǔn)算法均有提升,。
中圖分類號(hào): TP18
文獻(xiàn)標(biāo)識(shí)碼: A
DOI: 10.19358/j.issn.2096-5133.2021.04.005
引用格式: 李永強(qiáng),,白天. 基于卷積LSTM的視頻中Deepfake檢測(cè)方法[J].信息技術(shù)與網(wǎng)絡(luò)安全,2021,,40(4):28-32.
文獻(xiàn)標(biāo)識(shí)碼: A
DOI: 10.19358/j.issn.2096-5133.2021.04.005
引用格式: 李永強(qiáng),,白天. 基于卷積LSTM的視頻中Deepfake檢測(cè)方法[J].信息技術(shù)與網(wǎng)絡(luò)安全,2021,,40(4):28-32.
Deepfake detection method in videos based on convolutional LSTM
Li Yongqiang,,Bai Tian
(School of Software Engineering,University of Science and Technology of China,,Hefei 230026,,China)
Abstract: The face forgery technology represented by deepfake can replace the face in video with the target face by using a small amount of face data, so as to achieve the purpose of forgery video. The abuse of this kind of technology will bring adverse social effects, which need to be punished by using detection technology. The existing methods have some limitations, single frame detection algorithm ignores the dynamic defect of deepfake; when the data has defects, the model may fall into the trap of "learning specific face". In this paper, we propose a forgery face detection method in video, which uses the convolutional LSTM combined with CNN and LSTM to judge if a video is original or manipulated by deepfake. In addition, we propose a cutout method based on landmarks, which can inhibit the network from learning specific face. Experiments show that the accuracy of the baseline algorithm is improved in different scenes.
Key words : Deepfake detection;computer vision,;deep learning
0 引言
近年來,,基于深度學(xué)習(xí)技術(shù)的圖像生成技術(shù)迅速發(fā)展,視頻人臉偽造技術(shù)也隨之日趨成熟,。利用此類技術(shù)的人臉偽造技術(shù)已經(jīng)可以欺騙普通人類[1],。但這些技術(shù)的濫用也引發(fā)了一些社會(huì)問題,因?yàn)檫@些技術(shù)可以利用公眾人物公開的視頻,、圖像素材,,偽造公眾人物出場(chǎng)的虛假視頻,發(fā)布虛假的言論,,或偽造色情影片,破壞名譽(yù),。由于Deepfakes項(xiàng)目[2]的廣泛流傳,,這一類技術(shù)常被通稱為Deepfake。為了避免Deepfake技術(shù)的濫用,,許多研究團(tuán)體做出了卓越的貢獻(xiàn),。ROSSLER A等人發(fā)布了包含大量Deepfake數(shù)據(jù)的公開數(shù)據(jù)集FaceForensics++[1],以幫助研究人員研究檢測(cè)算法,。Facebook開展了DFDC(Deepfake Detection Challenge)比賽并公布了訓(xùn)練數(shù)據(jù)集[3],。
本文詳細(xì)內(nèi)容請(qǐng)下載:http://forexkbc.com/resource/share/2000003474
作者信息:
李永強(qiáng),白 天
(中國(guó)科學(xué)技術(shù)大學(xué) 軟件學(xué)院,,安徽 合肥230026)
此內(nèi)容為AET網(wǎng)站原創(chuàng),,未經(jīng)授權(quán)禁止轉(zhuǎn)載。