da/sec scientific talk on Biometrics

Topic: Human and Algorithm Detection of Manipulated Face Images

by Mathias Ibsen
D19/2.03a, April 18, 2024 (Thursday), 12.00 noon

Keywords — Digital forensics, face manipulation detection, human performance, algorithm performance, information fusion

Abstract

„Research indicates that face images manipulated digitally can compromise the security of automated verification systems, such as those used at international borders. In these instances and in many countries, an adversary can obtain an identity document with a manipulated face image, which can be used to get unauthorised access to a system. Both human observers and algorithms have difficulty identifying these manipulated face images. In this study, we evaluate the detection capabilities of both humans and algorithms on images that have been retouched, face-swapped, or morphed. Particularly, we investigate the joint detection performance of algorithms and humans in a differential detection scenario where both a suspected and trusted image are presented together. We introduce a conditional approach for detecting face image manipulations, where the human examiners‘ decisions are considered only when the algorithm is uncertain. Our findings show that, in general, the used algorithm outperforms human examiners on the used data and that combining human and algorithm decisions can improve detection accuracy.“