東華大學圖書館 |

語系: 繁體中文

說明(常見問題)

回圖書館首頁

手機版館藏查詢

登入

回首頁 到查詢結果 [ null ]

切換: 標籤 | MARC模式 | ISBD

Micro Signal Extraction for Surface-...

Liu, Runze,

FindBook

Google Book

Amazon

博客來

Micro Signal Extraction for Surface-Based Authentication and Deepfake Detection /

紀錄類型:	書目-電子資源 : Monograph/item
正題名/作者:	Micro Signal Extraction for Surface-Based Authentication and Deepfake Detection // Runze Liu.
作者:	Liu, Runze,
面頁冊數:	1 electronic resource (132 pages)
附註:	Source: Dissertations Abstracts International, Volume: 84-02, Section: B.
Contained By:	Dissertations Abstracts International84-02B.
標題:	Digitization. -
電子資源:	https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=29228763
ISBN:	9798841528357

Micro Signal Extraction for Surface-Based Authentication and Deepfake Detection /
Liu, Runze,

Micro Signal Extraction for Surface-Based Authentication and Deepfake Detection /Runze Liu. - 1 electronic resource (132 pages)

Source: Dissertations Abstracts International, Volume: 84-02, Section: B.

With the fast development of computer vision technology and the mobile phone industry, it becomes desirable to use camera-captured images for forensic applications, such as anti-counterfeiting of paper-based documents and even integrated circuits (IC) chips. On the other hand, the development of deepfake technology has raised public concern. This dissertation will discuss methods for paper/IC authentication and deepfake detection. The first part of this dissertation introduces authentication systems based on physical unclonable function (PUF). When using the mobile camera as capturing device for paper PUF, we develop two enhanced geometric reflection models that take into account the ambient light for norm map estimation. Also, new physical features, the high spatial frequency subbands, are proposed as discriminative features because they can better characterize the microstructures of paper surfaces. We also use flatbed scanners, which have a well-controlled experimental environment, to investigate key research questions in the paper surface-based authentication system. This dissertation studies the effect of specular reflection, paper patch size, blurring effect, and different physical features in the setup of flatbed scanners. The understanding from the results using scanner helps the design of paper surface-based authentication system in the real-world scenario. Like the paper surface, the microstructure of IC chip surfaces is also random. Therefore, the uniqueness of the surface of IC chips can be used for IC authentication. To the best of our knowledge, optical PUFs have not been used for IC chip authentication. We investigate into using camera/flatbed scanners to capture images of IC chip surfaces, and then obtain physical features of the chip surfaces for authentication. To make it more applicable in the real-world scenario, we also build fast verification schemes for IC chip authentication using a camera to capture videos of the chip surfaces.The second part of this dissertation studies and exploits the traces left by neural networks for deepfake detection. We first investigate an intuitive but understudied characteristic, "nonsmoothness", caused by max pooling and ReLU activation in the neural networks. We mathematically define the nonsmoothness and use synthetic data to confirm its existence. This dissertation also studies the statistical properties of nonsmoothness and models the events of nonsmoothness. The nonsmoothness can be regarded as a trace left by neural networks, which may be exploited for deepfake detection since high-quality deepfake videos are generated by neural networks. We propose the double operation method to exploit the traces left by neural network for deepfake detection. For a deepfake video already with traces left by neural network, the reconstructed version of the deepfake video also has traces. For a raw video without traces, the reconstructed version of it will contain traces left by neural network. Therefore, the videos before and after the reconstruction can be compared for deepfake detection. Siamese neural network has been used to build a classifier, and the detection results outperform state-of-the-art methods based on end-to-end CNN structures.

English

ISBN: 9798841528357Subjects--Topical Terms:

3681264
Digitization.

Micro Signal Extraction for Surface-Based Authentication and Deepfake Detection /
LDR:04686nmm a22003973i 4500 001 2437627
005 20251124054948.5
006 m o d
007 cr|nu||||||||
008 261011s2022 miu||||||m |||||||eng d
020 $a 9798841528357
035 $a (MiAaPQD)AAI29228763
035 $a (MiAaPQD)NCState_Univ18402039584
035 $a AAI29228763
040 $a MiAaPQD $b eng $c MiAaPQD $e rda
100 1 $a Liu, Runze, $e author. $3 3828680
245 1 0 $a Micro Signal Extraction for Surface-Based Authentication and Deepfake Detection / $c Runze Liu.
264 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2022
300 $a 1 electronic resource (132 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 84-02, Section: B.
500 $a Advisors: Lobaton, Edgar; Wu, Tianfu; Ghosh, Sujit; Dai, Huaiyu Committee members: Wong, Chau-Wai.
502 $b Ph.D. $c North Carolina State University $d 2022.
520 $a With the fast development of computer vision technology and the mobile phone industry, it becomes desirable to use camera-captured images for forensic applications, such as anti-counterfeiting of paper-based documents and even integrated circuits (IC) chips. On the other hand, the development of deepfake technology has raised public concern. This dissertation will discuss methods for paper/IC authentication and deepfake detection. The first part of this dissertation introduces authentication systems based on physical unclonable function (PUF). When using the mobile camera as capturing device for paper PUF, we develop two enhanced geometric reflection models that take into account the ambient light for norm map estimation. Also, new physical features, the high spatial frequency subbands, are proposed as discriminative features because they can better characterize the microstructures of paper surfaces. We also use flatbed scanners, which have a well-controlled experimental environment, to investigate key research questions in the paper surface-based authentication system. This dissertation studies the effect of specular reflection, paper patch size, blurring effect, and different physical features in the setup of flatbed scanners. The understanding from the results using scanner helps the design of paper surface-based authentication system in the real-world scenario. Like the paper surface, the microstructure of IC chip surfaces is also random. Therefore, the uniqueness of the surface of IC chips can be used for IC authentication. To the best of our knowledge, optical PUFs have not been used for IC chip authentication. We investigate into using camera/flatbed scanners to capture images of IC chip surfaces, and then obtain physical features of the chip surfaces for authentication. To make it more applicable in the real-world scenario, we also build fast verification schemes for IC chip authentication using a camera to capture videos of the chip surfaces.The second part of this dissertation studies and exploits the traces left by neural networks for deepfake detection. We first investigate an intuitive but understudied characteristic, "nonsmoothness", caused by max pooling and ReLU activation in the neural networks. We mathematically define the nonsmoothness and use synthetic data to confirm its existence. This dissertation also studies the statistical properties of nonsmoothness and models the events of nonsmoothness. The nonsmoothness can be regarded as a trace left by neural networks, which may be exploited for deepfake detection since high-quality deepfake videos are generated by neural networks. We propose the double operation method to exploit the traces left by neural network for deepfake detection. For a deepfake video already with traces left by neural network, the reconstructed version of the deepfake video also has traces. For a raw video without traces, the reconstructed version of it will contain traces left by neural network. Therefore, the videos before and after the reconstruction can be compared for deepfake detection. Siamese neural network has been used to build a classifier, and the detection results outperform state-of-the-art methods based on end-to-end CNN structures.
546 $a English
590 $a School code: 0155
650 4 $a Digitization. $3 3681264
650 4 $a Histograms. $3 3683681
650 4 $a Scanners. $3 3562953
650 4 $a Neural networks. $3 677449
690 $a 0800
710 2 $a North Carolina State University. $e degree granting institution. $3 3765716
720 1 $a Lobaton, Edgar $e degree supervisor.
720 1 $a Wu, Tianfu $e degree supervisor.
720 1 $a Ghosh, Sujit $e degree supervisor.
720 1 $a Dai, Huaiyu $e degree supervisor.
773 0 $t Dissertations Abstracts International $g 84-02B.
790 $a 0155
791 $a Ph.D.
792 $a 2022
856 4 0 $u https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=29228763