UTFaculteitenEEMCSDisciplines & departementenDMBResearchComputer Vision and BiometricsFace morphing - Detection of facial photograph manipulation by morphing

Face morphing - Detection of facial photograph manipulation by morphing

Project duration:

Jul 2018 - June 2024

Face morphing

Detection of facial photograph manipulation by morphing

Project summary:

Facial photo morphing is the combination of photographs of two persons in such a way that the resulting facial image looks like both persons. Photo morphing is a threat for the process of application of identity documents, because it offers the possibility for two different persons to use the same identity document. In 2014, with the publication ”The magic passport”, Ferrara et al demonstrated that automated facial recognition software can be fooled using morphed passport photographs, i.e. live images of both persons result in high comparison scores when compared to the morphed photograph. Recent investigations showed that also humans can be fooled if the morphs are of high quality as is shown by the research of the University of New South Wales and our own research.

Several attempts have been made to detect morphing by training classifiers using morphed and genuine photographs. However, it turned out that these methods rather detect traces of the specific morphing process rather than the morphing itself. n example are double compression artifacts that are caused by using jpeg compressed photographs to perform the morphing on and then compress the morphed photograph again. A classifier can be trained to detect these artifacts, but it will fail if uncompressed photographs are used as a source.. As to date, even if there exist several methods to detect the above mentioned artifacts, there are no reliable methods to detect high quality morphs.

The aim of this research project is to develop a deep knowledge of the morphing processes and methods to reliably detect if facial photographs are manipulated, in particular by morphing, by a more fundamental approach as opposed to the more data driven approaches used by other researchers.

Project Leader:

Funding:

RVIG