Jeremy Dawson（西弗吉尼亚大学）：CITeR 高质量变形数据集.pdf

1、Center for Identification Technology ResearchCITeRCITeR High-Quality Morph DatasetsNasser Nasrabadi(JHU),Jeremy Dawson(WVU),Chen Liu(CU),Stephanie Schuckers(CU),David Doermann(UB),Srirangaraj Setlur(UB),Siwei Lyu(UB)Center for Identification Technology ResearchCITeR2ProblemThere is a lack of high-qu

2、ality morph face image datasets that can be used to train detectors,evaluate performance,etc.Research Goals:To generate high-quality sequestered face morphing datasets(including manually check/touchup,print&scan,compressed,perturbed)that can benefit and serve as the milestone for developing defense

3、mechanisms against face morphing attacks(NIST FRVT Morph Detection Report,06/202023,https:/pages.nist.gov/frvt/html/frvt_morph.html)Generate a high-quality synthetic face database for morph face generation.Why is this important?Single and differential morph detectors are used to identify morphed fac

4、es during the passport applications and at the border-crossing for entry and exit scenario.2Center for Identification Technology ResearchCITeR3Our ApproachGenerate low-quality to high-quality morph images.Assemble several databases with different image resolutions to represent the faces captured dur

5、ing border crossing,webcam,and high-quality mugshot passport photos.Generate morph images using WVU,CU,and BU advanced morphing techniques(e.g.,combined land-mark and StyleGAN-based morphing,manipulating StyleGAN-based latent codes for morphing,transformer-based and diffusion-based morphing techniqu

6、es.Generate print&scan(P&S),compressed,perturbed morph images.Use several commercial printer&scanners to generate different quality P&S morph images.Generate compressed,adversarially perturbed,and manually touched up morph images.Evaluate the visual and quantitative quality of morph images.Generate