Iris Deidentification With High Visual Realism for Privacy Protection on Websites and Social Networks

Mauro Barni,Angelo Genovese,Vincenzo Piuri,Ruggero Donida Labati,Fabio Scotti

doi:10.1109/access.2021.3114588

Mauro Barni, Angelo Genovese + Show 3 more

Open Access

https://doi.org/10.1109/access.2021.3114588

Copy DOI

Abstract

The very high recognition accuracy of iris-based biometric systems and the increasing distribution of high-resolution personal images on websites and social media are creating privacy risks that users and the biometric community have not yet addressed properly. Biometric information contained in the iris region can be used to automatically recognize individuals even after several years, potentially enabling pervasive identification, recognition, and tracking of individuals without explicit consent. To address this issue, this paper presents two main contributions. First, we demonstrate, through practical examples, that the risk associated with iris-based identification by means of images collected from public websites and social media is real. Second, we propose an innovative method based on generative adversarial networks (GANs) that can automatically generate novel images with high visual realism, in which all the biometric information associated with an individual in the iris region has been removed and replaced. We tested the proposed method on an image dataset composed of high-resolution portrait images collected from the web. The results show that the generated deidentified images significantly reduce the privacy risks and, in most cases, are indistinguishable from real samples.

Highlights

T HE number of high-resolution images and videos uploaded by users on social networks and web-based applications is constantly increasing
To analyze the deidentification capability of the proposed method, we evaluated the accuracy of different biometric recognition schemes for real images and deidentified images, analyzed the matching scores obtained by matching real iris images and deidentified images, and evaluated the capability of the proposed generative adversarial networks (GANs) to generate random textures
We evaluated the accuracy of a neural network with a unified deep learning architecture (UNINET) [50], a method based on machine learning and binary statistical image features (BSIF) [51], and the following recognition methods implemented in the University of Salzburg Iris Toolkit (USIT) version 3.0 [8]: log Gabor (LG) [52], complex Gabor (CG) [23], local intensity variations (CR) [53], cumulative sums of grayscale blocks (KO) [54], and quadratic spline wavelet (QSW) [55]

Summary

Introduction

T HE number of high-resolution images and videos uploaded by users on social networks and web-based applications is constantly increasing. These images present a relevant privacy risk since biometric recognition could be performed by third parties without the explicit consent of the owners [1]. The need to protect high-resolution images posted on social media from the possibility of biometric recognition was proven in recent studies [2]. Recent iris recognition techniques introduce the possibility of performing biometric recognition by using portrait pictures uploaded on websites or social networks [6].

Methods

Results

Conclusion