基於視覺感知模型之深度偽造對抗性擾動;Adversarial Perturbation against Deepfakes based on Visual Perceptual Model

NCU Institutional Repository > 資訊電機學院 > 資訊工程研究所 > 博碩士論文 > Item 987654321/92969

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/92969

题名:	基於視覺感知模型之深度偽造對抗性擾動;Adversarial Perturbation against Deepfakes based on Visual Perceptual Model
作者:	王冠中;Wang, Kuan-Chung
贡献者:	資訊工程學系
关键词:	深度偽造;對抗性浮水印;深度學習;Deepfakes;adversarial watermark;deep learning
日期:	2023-01-11
上传时间:	2024-09-19 16:36:03 (UTC+8)
出版者:	國立中央大學
摘要:	深度偽造技術的出現對於數位視訊真實性帶來很大的威脅，近期許多研究針對深度偽造內容是否存在於視訊中發表相關的偵測與辨識方法，另也有研究學者提出在公開的影像中嵌入所謂對抗性浮水印，試圖使深偽模型所生成的竄改影像內容偏離預期結果，避免產生有效的竄改內容。現有的浮水印方法多於像素域中加入這種對抗性訊號，然而為了避免過強的浮水印訊號損及原影像畫質，無法在像素值施予較大幅度的改變。本研究提出於影像頻率域中嵌入對抗性浮水印，將影像轉換至亮度及色度空間後計算離散餘弦轉換(Discrete Cosine Transform, DCT)，透過Watson感知模型計算在不被人眼察覺下，確保DCT係數的修改低於可能的最大改變量，並依此決定浮水印在訓練階段時的修改步長。實驗結果顯示，所加入的高強度浮水印訊號確實能使深偽模型所生成的影像更容易發生嚴重失真，同時藉由計算影像畫質衡量來證實這樣的方法與像素值嵌入方法相比可有效降低對於原影像畫質的破壞。;The emergence of Deepfakes poses a serious threat to the authenticity of digital videos. Recently, many studies have proposed methods for detecting and identifying the presence of Deepfakes in videos. On the other hand, some researchers adopted the approach of digital watermarking by embedding adversarial signals in public images to make the tampering results generated by Deepfake models deviate from their expected goals, so as to avoid producing effective falsified content. Most existing watermarking methods embedded such adversarial signals in the pixel domain. However, in order to prevent the quality of original image from being damaged by overly strong watermark signals, making large changes to the pixel values is not feasible. In this research, we propose to embed the adversarial watermark signals in the frequency domain of images. After converting the image from RGB color channels to YUV channels, the DCT (Discrete Cosine Transform) is applied on each channel. The Watson’s perception model is employed to determine the maximum possible change of DCT coefficients to ensure that the modification won’t be noticed by the human eyes. The perceptual mask is also used to determine the modification step size of the watermark in the training stage. The experimental results show that embedding such stronger watermarking signals can introduce more severe distortions on the image generated by the Deepfake models.
显示于类别:	[資訊工程研究所] 博碩士論文

文件中的档案:

档案	描述	大小	格式	浏览次数
index.html		0Kb	HTML	44	检视/开启

在NCUIR中所有的数据项都受到原著作权保护.

社群 sharing

数据加载中.....