PRIVACY-PRESERVING FACE REDACTION USING CROWDSOURCING

Face redaction is used to deidentify images of people. Most approaches depend on face detection, but automated algorithms are still not adequate for sensitive applications in which even one unredacted face could lead to irreversible harm. Human annotators can potentially provide the most accurate detection, but only trusted annotators should be allowed to see the faces of privacy-sensitive applications. Redacting more images than trusted annotators could accommodate requires a new approach.

This dissertation leverages the characteristics of human perception of faces in median-filtered images in a human computation algorithm to engage crowd workers to redact faces—without revealing the identities. IntoFocus, a system I developed, permits robust face redaction with probabilistic privacy guarantees. The system's design builds on an experiment that measured the filter levels and conditions where participants could detect and identify faces.

Pterodactyl is a system that focuses on increasing the productivity of crowd-based face redaction systems. It uses the AdaptiveFocus filter, a filter that combines human perception of faces in median filtered images with a convolutional neural network to estimate a median filter level for each region of the image to allow the faces to be detected and prevent them from being identified.