Privacy-Preserving Face Detection: An Integrated Approach Using Multi-Task Cascaded Convolutional Networks and Randomized Convolutional Layers

Abstract

Facial analysis is still a major challenge in computer vision, but the extensive use of biometric systems has caused big privacy worries since facial images contain identity information which is sensitive and persistent. This study introduces a Privacy-Preserving AI (PP-AI) process that maintains face-detection capability while reducing visual leakage containing the identifying features to the extent possible. A novel method involving Multi-Task Cascaded Convolutional Network (MTCNN) for locating faces together with blurring using Gaussian of the detected area of Interest (AOI), then randomized convolutional feature extraction is done to remove identity signals while keeping characteristics needed for face recognition. The results of tests carried out on standardized benchmark datasets reveal that the face mask preserving CNN gets a detection accuracy rate of 95.2% - 97.4%, while the conventional one attains the accuracy rate of 99.1% - 99.8%. This equates to a 2-4% drop in accuracy, which in this study, is called the Privacy Tax. Accuracy metrics including precision recall as well as F1_score are still quite impressive, at the value of 0.94, 0.93, and 0.935 respectively. Results show that the privacy preserving method introduced in this paper can effectively remove identity-related visual signals while at the same time keeping the level of face detection performance high, a feature which would make it most of all appealing for privacy critical computer vision applications.