Balancing Privacy Preservation and Utility in Anti-Face Detection Systems

The advancement in facial detection has led to increased privacy concerns as it can potentially lead to ma-licious activities such as identity theft and illegal surveillance. Hence, several anti-face detection systems have been developed to address this concern. However, existing systems are trained end-to-end and do not give information as to how image features e.g. resolution, texture, and color can contribute to detection/misdetection. In this work, we find the significant image features in predicting the minimum Gaussian noise,
$sigma_{min}$
, necessary to avoid face detection. Three regression models were trained to determine the minimum level of perturbation. The multilayer perceptron (MLP) regression model exhibited the best performance, with 38.02 % of the detected faces being effectively concealed. The results from this study show that the image features such as local binary pattern, width, value, and height significantly impacted the predicted values of sigma, and can be used to accurately determine the minimum level of Gaussian noise to prevent face detection systems from detecting faces in images.