Development of a framework to evaluate and characterise Deep Learning models used for deep-fake detection in the identity fraud industry.

The identity fraud industry has been developing computer vision models based on Deep Learning techniques to identify deep-fake images of people's faces that could be used, for example, to open bank accounts. The work aims to identify image properties that are relevant to characterize the performance of the binary classifier, this information would provide valuable insights for the design and training of the models. In this work, we are following two research lines: inquiring about intrinsic image properties such as brightness, image quality, etc. And also the impact of specific features present in the images such as gender, ethnicity, and the presence of specific objects.