https://scholars.lib.ntu.edu.tw/handle/123456789/632448
標題: | Deepfake Algorithm Using Multiple Noise Modalities with Two-Branch Prediction Network | 作者: | Hsu H.-W JIAN-JIUN DING |
公開日期: | 2021 | 起(迄)頁: | 1662-1669 | 來源出版物: | 2021 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2021 - Proceedings | 摘要: | In this paper, we propose a facial manipulation detection method based on multiple image noise analysis modalities and a two-branch prediction network to separation different types of forgery artifacts. The proposed architecture reveals whether the input image can be decomposed into a blending of two images from different sources, and checks whether some patches of the input image are generated from a deep learning networks at the same time. We observe that most of the existing forgery detection work] only focuses on finding one of the blending or manipulation artifacts in the input image. As a result, this method provides an effective way for forgery detection by simultaneously checking the manipulation and blending artifacts. In addition, for use with different types of image noise analysis modalities, our method can find more robust detection features in the high-frequency domain compared with traditionally detection in the RGB domain, thereby obtaining better performance. Extensive experiments show that our method outperforms other existing forgery detection methods on detecting synthesized face image, no matter on detecting training dataset or on detecting unseen face manipulation techniques. © 2021 APSIPA. |
URI: | https://www.scopus.com/inward/record.uri?eid=2-s2.0-85126725003&partnerID=40&md5=acc838805626995290ccb302d9963698 https://scholars.lib.ntu.edu.tw/handle/123456789/632448 |
SDG/關鍵字: | Face recognition; Frequency domain analysis; Image analysis; Branch prediction; Detection features; Detection methods; Forgery detections; Image noise analysis; Input image; Learning network; Multiple image; Proposed architectures; Robust detection; Deep learning |
顯示於: | 電機工程學系 |
在 IR 系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。