The Science of Disguise

Technological advances have made digital cameras ubiquitous, to the point where it is difficult to purchase even a mobile phone without one. Coupled with similar advances in face recognition technology, we are seeing a marked increase in the use of biometrics, such as face recognition, to identify individuals. However, remaining unrecognized in an era of ubiquitous camera surveillance remains desirable to some citizens, notably those concerned with privacy. Since biometrics are an intrinsic part of a person's identity, it may be that the only means of evading detection is through disguise. We have created a comprehensive database of high-quality imagery that will allow us to explore the effectiveness of disguise as an approach to avoiding unwanted recognition. Using this database, we have evaluated the performance of a variety of automated machine-based face recognition algorithms on disguised faces. Our data-driven analysis finds that for the sample population contained in our database: (1) disguise is effective; (2) there are significant performance differences between individuals and demographic groups; and (3) elements including coverage, contrast, and disguise combination are determinative factors in the success or failure of face recognition algorithms on an image. In this dissertation, we examine the present-day uses of face recognition and their interplay with privacy concerns. We sketch the capabilities of a new database of facial imagery, unique both in the diversity of the imaged population, and in the diversity and consistency of disguises applied to each subject. We provide an analysis of disguise performance based on both a highly-rated commercial face recognition system and an open-source algorithm available to the FR community. Finally, we put forth hypothetical models for these results, and provide insights into the types of disguises that are the most effective at defeating facial recognition for various demographic populations. As cameras become more sophisticated and algorithms become more advanced, disguise may become less effective. For security professionals, this is a laudable outcome; privacy advocates will certainly feel differently.

Author dc:creator

• Gibson, Katherine

Advisor dc:contributor.advisor

• Jonathan M. Smith

Statement dc:rights

• Katherine Gibson

Language dc:language

en

Repository record dc:identifier.uri

https://repository.upenn.edu/handle/20.500.14332/28578

OAI identifier oai:identifier

oai:repository.upenn.edu:20.500.14332/28578