An Assessment of Image-Cloaking Techniques Against Automated Face Recognition for Biometric Privacy

Abstract

Over the past two decades, Americans have aggressively increased the amount of facial data uploaded to the internet primarily via social media. This data is largely unprotected due to the dire lack of existing regulations protecting users from large scale face recognition in the United States, where the value of data trade is in the tens of billions. In its current state, facial privacy in the United States depends on American corporations opting not to collect the public data, an option rarely chosen. Much research has been done in the area of suppressing recognition abilities, giving users the ability to protect themselves. In our experiment, two techniques made publicly available: the Fawkes and LowKey algorithms are evaluated on their effectiveness in suppressing identification rates when applied to personal images. Through detailed assessment of match characteristic data, match score distributions, and image observations, we find that each algorithm performs where the other falls short both in identification suppression and preservation of the original image. The achieved results reveal a plethora of use cases for the currently available technology in addition to the reality that image cloaking techniques targeting social media use, both present and future, face a strict constraint of preserving image quality to the human eye while achieving enough perturbation to measurably increase users’ privacy.