Conditional Identity Disentanglement for Differential Face Morph Detection

TL;DR

This paper introduces a conditional generative network approach for detecting face morph attacks in identification documents, effectively disentangling identities and outperforming existing methods in cross-dataset scenarios.

Contribution

The novel use of cGAN for differential face morph detection enables identity disentanglement and information recovery, improving robustness across datasets and attack types.

Findings

Achieved 3% BPCER @ 10% APCER on intra-dataset evaluation.

Outperformed state-of-the-art with 4.6% BPCER @ 10% APCER on cross-dataset detection.

Demonstrated effectiveness on multiple face morph datasets.

Abstract

We present the task of differential face morph attack detection using a conditional generative network (cGAN). To determine whether a face image in an identification document, such as a passport, is morphed or not, we propose an algorithm that learns to implicitly disentangle identities from the morphed image conditioned on the trusted reference image using the cGAN. Furthermore, the proposed method can also recover some underlying information about the second subject used in generating the morph. We performed experiments on AMSL face morph, MorGAN, and EMorGAN datasets to demonstrate the effectiveness of the proposed method. We also conducted cross-dataset and cross-attack detection experiments. We obtained promising results of 3% BPCER @ 10% APCER on intra-dataset evaluation, which is comparable to existing methods; and 4.6% BPCER @ 10% APCER on cross-dataset evaluation, which outperforms state-of-the-art methods by at least 13.9%.