Privacy-Preserving Face Matching Using Frequency Components

Abstract

Faces constitute a susceptible portion of an image, which in this age has become vulnerable with the increasing use of applications involving face detectors. Increasing concerns over the privacy of people have led to research involving anonymizing faces, privacy-preserving feature selection from facial images, active learning to obfuscate the visual appearance of a detected face, and many other approaches that make the face unviewable during its use in an application, specifically in the case of facial matching or face recognition. Along the lines of preserving privacy of detected faces during a face matching process, we aim to provide a possible solution to the privacy issues by utilizing the frequency components of an image instead of using the intensity components. This process involves computing a frequency transform of the detected facial image and using the frequency coefficients in the matching process. The matching is established using metrics such as the cosine distance, correlation value, and the Manhattan distance between two facial images. We use the three metrics to compute a combined score and establish a valid match by use of empirically derived threshold values. We analyze the role of specific DCT coefficients in making the combined decision and test the overall algorithm on the Aberdeen and Utrecht face image dataset.