Hue-Tiful Privacy: Impact of Color Correction on Privacy Preserving Techniques

In light of growing concern surrounding the unauthorized collection of online photos without consent, privacy-minded individuals are seeking more robust methods to safeguard their data. One promising solution for preserving privacy is the use of tools like Fawkes. Such tools attempt to shield individuals from unwanted face recognition by implementing pixel-level alterations, a de-identification technique commonly referred to as “cloaking”. We explore the potential enhancements in privacy preservation by introducing a novel approach: applying color correction to images prior to the cloaking process. Our goal is to determine if this additional step could bolster individuals' privacy. To assess the effectiveness of color correction in conjunction with cloaking, we employ three face recognition models: ArcFace, AdaFace, and MagFace. Using facial features extracted from cloaked images with and without color correction, we leveraged metrics such as d-prime and Earth Mover's Distance to compare the score distributions of non-color-corrected cloaked images with those that underwent color-correction before cloaking. Our findings show that color correction augments the protection provided by the cloaking process.