Practical Evaluation of Face Morphing Attack Detection Methods

Abstract

AbstractFace morphing is a technique to combine facial images of two (or more) subjects such that the result resembles both subjects. In a morphing attack, this is exploited by, e.g., applying for a passport with the morphed image. Both subjects who contributed to the morphed image can then travel using this passport. Many state-of-the-art face recognition systems are vulnerable to morphing attacks. Morphing attack detection (MAD) methods are developed to mitigate this threat. MAD methods published in literature are often trained on a limited number of or even a single dataset where all morphed faces are created using the same procedure. The resulting MAD methods work well for these specific datasets, with reported detection rates of over 99%, but their performance collapses for face morphs created using other procedures. Often even simple image manipulations, like adding noise or smoothing cause a serious degradation in performance of the MAD methods. In addition, more advanced tools exist to manipulate the face morphs, like manual retouching or morphing artifacts can be concealed by printing and scanning a photograph (as used in the passport application process in many countries). Furthermore, datasets for training and testing MAD methods are often created by morphing images from arbitrary subjects including even male-female morphs and morphs between subjects with different skin color. Although this may result in a large number of morphed faces, the created morphs are often not convincing and certainly don’t represent a best effort attack by a criminal. A far more realistic attack would include careful selection of subjects that look alike and create high quality morphs from images of these subjects using careful (manual) post-processing. In this chapter we therefore argue that for robust evaluation of MAD methods, we require datasets with morphed images created using a large number of different morphing methods, including various ways to conceal the morphing artifacts by, e.g., adding noise, smoothing, printing and scanning, various ways of pre- and post-processing, careful selection of the subjects and multiple facial datasets. We also show the sensitivity of various MAD methods to the mentioned variations and the effect of training MAD methods on multiple datasets.