Hanoi university of science and technology Master thesis Facial image forgery and detection: benchmark and application Pham Minh Tam Tam.vn School of Information and Cammunication Technology Supervisor: Asse. Huynh Quyct Thang Supervisor's signaLure Institution: school of Information ard Communicetion Tectuology Hanoi, 12/2021 Declaration of Authorship J, Pham Minh ‘Iam, declare that this caes's titled, "Lacial image forgery and detection: benchmark and application” and the work presented in it are my own. { confirm that: » ‘This work was done wholly ar mainly while in candidature for a research degree at this University. ¢ Where | have consulted the publisted work of others, this is always clearly attributed.
© Where I have quoted from the work of others, the source is always given. With the exreption of suca quotations, this thesis is entirely may own work. Student Siguatare sed Name Acknowledgements J would like to extend thanks to the many people, wao so generously contributed to the work presented in this thesis confirmation report. T would fest like to thank any supervisur, Prof, Huynh Quyet Thang xLo give me su aany wonderfal apportumities and sive me so many valiable guidance thronghont my at J would like to thank Dr.
Menry Nguyen and Dr, Muynh Thanh ‘rung who provided me wilh Lie looks that, Taveded te chuuse the right direcuiou and successfully complete ay dissertation, ‘Taeir insightful feadhack pushed me to sharpen my thirking and bronght. my work to a higher level I also like to thank my family, my girl friend and my friends for their wise counsel and sympathetic ear. You are always there for re. Finally, T wenld Eke to thank Vingronn Tnnovation Foundation (VINTF) for the financial support, List of Tables 1.1 Comparison Between Existing Benehmarks on Faeial Forensis .1 Taxonomy of visual forgery teehniqMes.1 Statistie of real datasets.1 Model size and detection speed .2 Statistic train, validate, test of existing đatasels.3 Performance (Aecuracy|Precision|Reeall|F score) of visual forensic techniques on UNE REAR ce woe ranh nhé nà ên Hà nà và sendDneea 06g sử số 48 5.4 Statistie train, validate, test of synthetie datasets .5 Performance of visual forensics techniques against visual forgery techniques .6 Performance guideline for visual forensics ©.
ee 59 List of Figures 21 `. 22 Đbnwdltllnn ợNN bú n0 2/2001 0% ch tú THE be pà eNI606/8165 10 Díh trà ga Max and average pooling 2 Global average pooling. 25 Flatten and fully connected layer. vị bu co co 2202206002 0002 xà cóc Overview of Efficient-Froqueney pipeline.
uc dc co cu cu cọ nu ng chu nà nà hà 33 TWÁDĐTHHẰNN sa hà nung 2a ổn cú VU we wa mare Ube gio tá TH da NEEINEUNÊN cone nàooaninsoonhatrgcvncgin bạn danh HH đi, 35001016 35 ` "5.NNNaN-N- HH1ẽă HH 41 Dual benchmarking amework. 40 42 se OF ficial i6sty dat. cà cà 6à ÍỐ co Go cÍŸŸc G ÍÍ ko cc 4l Images from DBD dataset. eee eee 42 51 Generalization ability of forensie teehniques.-‹+ 50 52 Overlapping feature of forgery teehniques.‹ c- 5l 53 Suspicious region of forged images.
cà cốc cà CONTENTS Abstract: 1 Introduction Lit l0 "tt on Sea Mer twas $9.58 OS WSEAS 12 Research Problems i. cise as aa sie sew oe oo KỆ CðHRIBHHUBE- dạ cá cà nu nàng nọ Hà eee nut LA TheawOutln .‹ 2 Preliminaries and literature survey 2A Image classification problem.2 Visual forgery techniques .1 Graphics-based teehniqhe.2 Feature-based teehniqns .- 23 3⁄3 Visual forensics teehmique.1 Computer vision tefhnique .2 Deep learning technique. 38 3 Proposed facial forgery detection models 31 3.1 Efficient-Frequeney model 0.2 WADD (Wavelet Attention for deepfake detection) model a Abstract In recent years, visual forgery has reached a level of sophistication that aumars cannot iden- tify <taud, which poses a significant threat to information security. A wide range of malicious applications lave emerged, suc as fake uews, defamation or Llackuailing of velcbrivies, inper~ suuation uf politie‘ans in polilical warfare, aud the spreading of rumours lo aturact views.
rich hady of visnal favensir technicnes has been propored in an attempt to stop this dangerous trend. In this thesis, | introduce nvo new models with name Kfficient-Hrequency and WADD to improve the result of fake face images detection problem, I also present a benchmark Uhat provides iuCepth insights inlo visuel Jurgery and visual loreneies, using a comprelwusive aud empirical approach and propose a. novel end-to-end visnalforensic framework that: car corporate different modalit'es toefficiently classify real and forged enntenis. Mara specifica'ly, we develop an independent framework Lal inlegrates slale-of-Llearts counterfeit generators and delectons, aud sucasure Lhe perforinauc of (hee lechuiques using various crileria, We also perform au exhaustive analysis of the benchmarking results, to determine the characterietics of the methods that serve as a.
comparative reference in this never-ending war between measures and countermeasures. Simdent: ature and Name CONTENTS Abstract: 1 Introduction Lit l0 "tt on Sea Mer twas $9.58 OS WSEAS 12 Research Problems i. cise as aa sie sew oe oo KỆ CðHRIBHHUBE- dạ cá cà nu nàng nọ Hà eee nut LA TheawOutln .‹ 2 Preliminaries and literature survey 2A Image classification problem.2 Visual forgery techniques .1 Graphics-based teehniqhe.2 Feature-based teehniqns .- 23 3⁄3 Visual forensics teehmique.1 Computer vision tefhnique .2 Deep learning technique. 38 3 Proposed facial forgery detection models 31 3.1 Efficient-Frequeney model 0.2 WADD (Wavelet Attention for deepfake detection) model a List of Figures 21 `.
22 Đbnwdltllnn ợNN bú n0 2/2001 0% ch tú THE be pà eNI606/8165 10 Díh trà ga Max and average pooling 2 Global average pooling. 25 Flatten and fully connected layer. vị bu co co 2202206002 0002 xà cóc Overview of Efficient-Froqueney pipeline. uc dc co cu cu cọ nu ng chu nà nà hà 33 TWÁDĐTHHẰNN sa hà nung 2a ổn cú VU we wa mare Ube gio tá TH da NEEINEUNÊN cone nàooaninsoonhatrgcvncgin bạn danh HH đi, 35001016 35 ` "5.NNNaN-N- HH1ẽă HH 41 Dual benchmarking amework.
40 42 se OF ficial i6sty dat. cà cà 6à ÍỐ co Go cÍŸŸc G ÍÍ ko cc 4l Images from DBD dataset. eee eee 42 51 Generalization ability of forensie teehniques.-‹+ 50 52 Overlapping feature of forgery teehniques.‹ c- 5l 53 Suspicious region of forged images. cà cốc cà 54 Eifects of illumination factors.
{co 55 Robustness against noises.6 Robustness against image resolution 57 Influence of missing information. 58 Adaptivity to image compression CONTENTS Abstract: 1 Introduction Lit l0 "tt on Sea Mer twas $9.58 OS WSEAS 12 Research Problems i. cise as aa sie sew oe oo KỆ CðHRIBHHUBE- dạ cá cà nu nàng nọ Hà eee nut LA TheawOutln .‹ 2 Preliminaries and literature survey 2A Image classification problem.2 Visual forgery techniques .1 Graphics-based teehniqhe.2 Feature-based teehniqns .- 23 3⁄3 Visual forensics teehmique.1 Computer vision tefhnique .2 Deep learning technique. 38 3 Proposed facial forgery detection models 31 3.1 Efficient-Frequeney model 0.2 WADD (Wavelet Attention for deepfake detection) model a Abstract In recent years, visual forgery has reached a level of sophistication that aumars cannot iden- tify <taud, which poses a significant threat to information security.
A wide range of malicious applications lave emerged, suc as fake uews, defamation or Llackuailing of velcbrivies, inper~ suuation uf politie‘ans in polilical warfare, aud the spreading of rumours lo aturact views. rich hady of visnal favensir technicnes has been propored in an attempt to stop this dangerous trend. In this thesis, | introduce nvo new models with name Kfficient-Hrequency and WADD to improve the result of fake face images detection problem, I also present a benchmark Uhat provides iuCepth insights inlo visuel Jurgery and visual loreneies, using a comprelwusive aud empirical approach and propose a. novel end-to-end visnalforensic framework that: car corporate different modalit'es toefficiently classify real and forged enntenis.
Mara specifica'ly, we develop an independent framework Lal inlegrates slale-of-Llearts counterfeit generators and delectons, aud sucasure Lhe perforinauc of (hee lechuiques using various crileria, We also perform au exhaustive analysis of the benchmarking results, to determine the characterietics of the methods that serve as a. comparative reference in this never-ending war between measures and countermeasures. Simdent: ature and Name 4 Proposed dual benchmarking framework for facial forgery and detection tech- niques 39 41 EWAmeWGĂ. co họ en eee eee emt ee ee ne 42 4.
ete eet eee nee 43 Mecsuremenbts te có 209096/3V5: te Hh HR HS OPV RR TS Be HE wR d4 Experimental Proeedures. eee eee 5 Experimental Results and Performance Analysis 46 5A Bificletiey CAMBAERIOD ona sia.ane ae ae von vou whe wo imo 0á 0k nh mw 46 5.2 End-to-end comparison with existing datasets. ee a7 58 Dual-benchmarking compatison 0. 000 cee cece eee eee 48 5.1 Forensic generalisation and forgery feature overlappingồ.2 Qualitative study of forensic-forgery duel 2.0 eee ee ñ1 Influence of contrast 606.0 ee eee eee eee 52 Sd Bifecte of brightnese ¿c2 ::c: cị cóc: cá 222032 2c cà Oe a 53 59.
ñ4 5⁄36 Robustness against image resolution .7 Influence of miwing iôrmation .-c<s- 56 5/38 - Adaptivity to image omprewion. có co c2 a7 5A ROHR AMMA iia cee Ba eo fn HO SS Seca nes bb đâu số xế aT 6 Conclusions 60 Reference 61 List of Tables 1.1 Comparison Between Existing Benehmarks on Faeial Forensis .1 Taxonomy of visual forgery teehniqMes.1 Statistie of real datasets.1 Model size and detection speed .2 Statistic train, validate, test of existing đatasels.3 Performance (Aecuracy|Precision|Reeall|F score) of visual forensic techniques on UNE REAR ce woe ranh nhé nà ên Hà nà và sendDneea 06g sử số 48 5.4 Statistie train, validate, test of synthetie datasets .5 Performance of visual forensics techniques against visual forgery techniques .6 Performance guideline for visual forensics ©. ee 59 CONTENTS Abstract: 1 Introduction Lit l0 "tt on Sea Mer twas $9.58 OS WSEAS 12 Research Problems i. cise as aa sie sew oe oo KỆ CðHRIBHHUBE- dạ cá cà nu nàng nọ Hà eee nut LA TheawOutln .‹ 2 Preliminaries and literature survey 2A Image classification problem.2 Visual forgery techniques .1 Graphics-based teehniqhe.2 Feature-based teehniqns .- 23 3⁄3 Visual forensics teehmique.1 Computer vision tefhnique .2 Deep learning technique.
38 3 Proposed facial forgery detection models 31 3.1 Efficient-Frequeney model 0.2 WADD (Wavelet Attention for deepfake detection) model a Chapter 1 Introduction 1.1 Context Forgery inmages are syatheliv inages whic are crvaled Uy touis iz computer. The eaten of these images does not ref ect the truth, misleading viewers about reality. Face is the most faked component due to the importance o? the face in determining a person’s identity, face spoolirg is the most common method to falsify the information of a photo. Tn fake face images.
a person in an existing image or video is replaced with sameane else's likenese ry image" have emerged in the recent: years, hecanse there ‘nave bean a lot of tools helping charge the content of images ench as Photoshap software. Although hese Wools are so powerful, users must Have a huge a mauunt of knowledge Lo use thet. As a rusulls, Lhe aumber of fake inaages is low aud it wakes @ lot of resources to make a lake photo. Bút nơy, rather than images simply being alvered by editing sofmware such as Photoshop or videos being deceptively edited.
there's a new breed of marhine-made fakes — and they could evertnally make it impossible for us to tell fact from fiction. With the development of deep learning tecaniques, there are aot of methods whick can generate "fake images!" quickly, easily and in bulk. "Deep fakes* are the most prominent form af what's being called “syathet'c media”: images sound and video that appear to have been cicated through traditional means but that Lavo, in fact. been woustructed by couples software, Deep fakes buve beer around Zor years aud, evea though Uheir mos! common use ty date has becu trarsplanling Une Leads of eclcbrilies onto the bodies of aclurs in poruogcaphic videus, they have the polential to create couvincing fuolage of any person doing anything, anywhere.