Detecting Morphed Face Attacks Using Residual Noise from Deep Multi-scale Context Aggregation Network

Abstract

<p> Along with the deployment of the Face Recognition Systems</p> <p>(FRS), concerns were raised related to the vulnerability</p> <p>of those systems towards various attacks including morphed</p> <p>attacks. The morphed face attack involves two different</p> <p>face images in order to obtain via a morphing process</p> <p>a resulting attack image, which is sufficiently similar</p> <p>to both contributing data subjects. The obtained morphed</p> <p>image can successfully be verified against both subjects visually</p> <p>(by a human expert) and by a commercial FRS. The</p> <p>face morphing attack poses a severe security risk to the</p> <p>e-passport issuance process and to applications like border</p> <p>control, unless such attacks are detected and mitigated.</p> <p>In this work, we propose a new method to reliably detect</p> <p>a morphed face attack using a newly designed denoising</p> <p>framework. To this end, we design and introduce a new</p> <p>deep Multi-scale Context Aggregation Network (MS-CAN)</p> <p>to obtain denoised images, which is subsequently used to</p> <p>determine if an image is morphed or not. Extensive experiments</p> <p>are carried out on three different morphed face image</p> <p>datasets. The Morphing Attack Detection (MAD) performance</p> <p>of the proposed method is also benchmarked against</p> <p>14 different state-of-the-art techniques using the ISO-IEC</p> <p>30107-3 evaluation metrics. Based on the obtained quantitative</p> <p>results, the proposed method has indicated the best</p> <p>performance on all three datasets and also on cross-dataset</p> <p>experiments.</p>

Face Morphing Attack Generation and Detection: A Comprehensive Survey

“…The first work in this area was introduced in [123] based on CNN-based denoising on colour channels. Furthermore, the residual noise is effectively captured using the deep CNN approach [122]. The use of residual noise has shown considerably good performance in terms of generalisation capabilities across different digital datasets.…”

“…The first work in this direction was based on using pre-trained networks such as AlexNet and VGG18, in which the features are fused and classified to detect a morphing attack [93]. Following this, several deep CNN pre-trained networks, such as AlexNet, VGG19, VGG-Face16, GoogleNet, ResNet18, ResNet150, ResNet50, VGG-Face2 and OpenFace [52], [43], [122], [112], [107], [108], [113], [71], [112], have been explored. Although deep CNNs have shown better performance than hand-crafted texture descriptor-based MAD methods on both digital and print-scan data, the generalisation capability of these approaches is limited across different print and scan datasets [110].…”

Morphing Attack Detection-Database, Evaluation Platform, and Benchmarking

“…Another approach to detecting morphing attacks was proposed by extracting the features from the "Photo Response Non-Uniformity" where the characteristics of the image sensor were employed to determine, if the image was morphed or not [12]. Motivated by the effectiveness of the noise modelling, better performing algorithms have been reported where the color space has been investigated to seek for residuals of the morphing process [36] including dedicated context aggregation networks to automatically model the noise [37].…”

“…• Need for cross-dataset evaluation: As different works have used in-house datasets generated using different approaches, the proposed methods are only evaluated on limited sets. Despite the proposed MAD approaches performing very well on the in-house datasets, no works have attempted to study the generalizable detection performance except in recent works [33], [37] which attempts to study the cross-dataset evaluation. The missing aspect from different studies suffer from validation of SOTA proposed approaches in terms of generalizable detection performance and also indicating the directions for future works.…”