- KonferenzbeitragAn anomaly detection approach for backdoored neural networks: face recognition as a case study(BIOSIG 2022, 2022) Alexander Unnervik and Sébastien Marcel; Brömme, Arslan; Damer, Naser; Gomez-Barrero, Marta; Raja, Kiran; Rathgeb, Christian; Sequeira Ana F.; Todisco, Massimiliano; Uhl, AndreasBackdoor attacks allow an attacker to embed functionality jeopardizing proper behavior of any algorithm, machine learning or not. This hidden functionality can remain inactive for normal use of the algorithm until activated by the attacker. Given how stealthy backdoor attacks are, consequences of these backdoors could be disastrous if such networks were to be deployed for applications as critical as border or access control. In this paper, we propose a novel backdoored network detection method based on the principle of anomaly detection, involving access to the clean part of the training data and the trained network.We highlight its promising potential when considering various triggers, locations and identity pairs, without the need to make any assumptions on the nature of the backdoor and its setup. We test our method on a novel dataset of backdoored networks and report detectability results with perfect scores.
- KonferenzbeitragFingervein Sample Image Quality Assessment using Natural Scene Statistics(BIOSIG 2022, 2022) Oliver Remy, Jutta Hämmerle-Uhl and Andreas Uhl; Brömme, Arslan; Damer, Naser; Gomez-Barrero, Marta; Raja, Kiran; Rathgeb, Christian; Sequeira Ana F.; Todisco, Massimiliano; Uhl, AndreasNatural Scene Statistics as used in non-reference image quality measures are proposed to be used as fingervein sample quality indicators. While NIQE and BRISQUE trained on common images with usual distortions do not work well in the fingervein quality context, their variants being trained on high and low quality fingervein sample data behave as expected from a biometric quality estimator. Experiments involve two publicly available fingervein datasets and two distinct template representations. The proposed (trained) quality measures are compared to a set of classical fingervein quality metrics which underlines their highly promising behaviour.
- KonferenzbeitragWorst-Case Morphs: a Theoretical and a Practical Approach(BIOSIG 2022, 2022) Una Kelly, Luuk Spreeuwers and Raymond Veldhuis; Brömme, Arslan; Damer, Naser; Gomez-Barrero, Marta; Raja, Kiran; Rathgeb, Christian; Sequeira Ana F.; Todisco, Massimiliano; Uhl, AndreasFace Recognition (FR) systems have been shown to be vulnerable to morphing attacks. We examine exactly how challenging morphs can become. By showing a worst-case construction in the embedding space of an FR system and using a mapping from embedding space back to image space we generate images that show that this theoretical upper bound can be approximated if the FR system is known. The resulting morphs can also succesfully fool unseen FR systems and are useful for exploring and understanding the weaknesses of FR systems. Our method contributes to gaining more insight into the vulnerability of FR systems.
- KonferenzbeitragOn the detection of morphing attacks generated by GANs(BIOSIG 2022, 2022) Laurent Colbois and Sébastien Marcel; Brömme, Arslan; Damer, Naser; Gomez-Barrero, Marta; Raja, Kiran; Rathgeb, Christian; Sequeira Ana F.; Todisco, Massimiliano; Uhl, AndreasRecent works have demonstrated the feasibility of GAN-based morphing attacks that reach similar success rates as more traditional landmark-based methods. This new type of “deep” morphs might require the development of new adequate detectors to protect face recognition systems. We explore simple deep morph detection baselines based on spectral features and LBP histograms features, as well as on CNN models, both in the intra-dataset and cross-dataset case. We observe that simple LBP-based systems are already quite accurate in the intra-dataset setting, but struggle with generalization, a phenomenon that is partially mitigated by fusing together several of those systems at score-level.We conclude that a pretrained ResNet effective for GAN image detection is the most effective overall, reaching close to perfect accuracy. We note however that LBP-based systems maintain a level of interest : additionally to their lower computational requirements and increased interpretability with respect to CNNs, LBP+ResNet fusions sometimes also showcase increased performance versus ResNet-only, hinting that LBP-based systems can focus on meaningful signal that is not necessarily picked up by the CNN detector.
- KonferenzbeitragTemplate Protection: On the need to adapt the current Unlinkability Evaluation Protocol(BIOSIG 2022, 2022) Simon Kirchgasser and Andreas Uhl; Brömme, Arslan; Damer, Naser; Gomez-Barrero, Marta; Raja, Kiran; Rathgeb, Christian; Sequeira Ana F.; Todisco, Massimiliano; Uhl, AndreasUsing ISO/IEC Standards an evaluation protocol exists which properties need to be fulfilled by each template protection scheme. However, in these standards it is not defined how a sensible and sensitive key selection should be done such that the demanded properties are reached. For the analysis regarding the ISO/IEC Standards properties only arbitrary, randomly selected keys are usually used, but it is not considered that in this case the key selection might result in insufficiently protected biometric images (templates). By the performed experiments using not only randomly selected keys, but also considering the best of insufficient (improper) keys, it was revealed that the unlinkability evaluation protocol is influenced by the key selection. Hence, it is recommended to use both mentioned key types to analyse template protection methods. This means that the protocol has to be fine-tuned, in so far that for the best of the worst key choice (which then has to be used in the protocol) unlinkability still needs to be given.
- KonferenzbeitragDeDiM: De-identification using a diffusion model(BIOSIG 2022, 2022) Hidetsugu Uchida, Narishige Abe and Shigefumi Yamada; Brömme, Arslan; Damer, Naser; Gomez-Barrero, Marta; Raja, Kiran; Rathgeb, Christian; Sequeira Ana F.; Todisco, Massimiliano; Uhl, AndreasAs a countermeasure against malicious authentication in a face recognition system using a face image obtained from SNS or the like, de-identification methods based on adversarial example have been studied. However, since adversarial example directly uses the gradient information of a face recognition model, it is highly dependent on the model, and a de-identification effect and image quality are difficult to achieve for an unknown recognition model. In this study, we propose a novel de-identification method based on a diffusion model, which has high generalizability to an unknown recognition model by applying minute changes to face shapes. Experiments using LFW showed that the proposed method has a higher de-identification effect for unknown models and better image quality than a conventional method using adversarial example.
- KonferenzbeitragWhen Facial Recognition Systems become Presentation Attack Detectors(BIOSIG 2022, 2022) Lazaro Janier Gonzalez-Soler, Kevin Abadi Barhaugen; Brömme, Arslan; Damer, Naser; Gomez-Barrero, Marta; Raja, Kiran; Rathgeb, Christian; Sequeira Ana F.; Todisco, Massimiliano; Uhl, AndreasRecently, biometric systems (BSs) have experienced a broad development mainly due to the great success of deep learning approaches. Generally, most BS provide high security and efficiency. However, they are still vulnerable to attack presentations (APs). To overcome such security issues, these schemes include a Presentation Attack Detection (PAD) module which determines whether the input sample stems from an AP or a bona fide presentation (BP). Traditionally, most PAD subsystems assess the biometric sample prior to the recognition module. In this work, we evaluate to what extent the inverted combination, where the biometric recognition module filters samples prior to the assessment of a PAD mechanism, leads to an overall PAD performance improvement. The experimental evaluation conducted over two well-known databases including challenging attacks, reports a significant improvement in the detection performance when input samples were first filtered by the biometric recognition: only 1% of the APs are accepted while at most 5% BPs are rejected by the PAD subsystem.
- KonferenzbeitragProcessing Information Extracted from the Human Body: Measurements of Biological and Behavioural Signals as a Unifying Link(BIOSIG 2022, 2022) Lydia Belkadi; Brömme, Arslan; Damer, Naser; Gomez-Barrero, Marta; Raja, Kiran; Rathgeb, Christian; Sequeira Ana F.; Todisco, Massimiliano; Uhl, AndreasThis contribution proposes the concept of “measurements of biological and behavioural signals” for interdisciplinary research on the automated processing of information from the human body. This concept has merits in mitigating legal definitions’ instability. We further aim to bridge legal and technical vocabularies, both responding to specific methodologies. Revising this concept should enable a more coherent approach and account for information about the human body, emerging sensing devices, and automated systems.
- KonferenzbeitragBIOSIG 2022 - Komplettband(BIOSIG 2022, 2022) Brömme, Arslan; Damer, Naser; Gomez-Barrero, Marta; Raja, Kiran; Rathgeb, Christian; Sequeira Ana F.; Todisco, Massimiliano; Uhl, Andreas
- KonferenzbeitragEvaluating Face Image Quality Score Fusion for Modern Deep Learning Models(BIOSIG 2022, 2022) Schlett, Torsten; Rathgeb, Christian; Tapia, Juan E.; Busch, Christoph; Brömme, Arslan; Damer, Naser; Gomez-Barrero, Marta; Raja, Kiran; Rathgeb, Christian; Sequeira Ana F.; Todisco, Massimiliano; Uhl, AndreasFace image quality assessment algorithms attempt to estimate the utility of face images for biometric systems, typically face recognition, since the performance of these systems can be limited by the image quality. Hand-crafted quality score fusion has previously been examined for a variety of mostly factor-specific quality assessment algorithms. This paper instead examines score fusion for various recent “monolithic” quality assessment deep learning models. The evaluation methodology is based on Error-versus-Reject-Characteristic partial-Area-Under-Curve values, which are used to quantitatively rank quality assessment configurations in a face recognition context. Mean quality score fusion configurations were found to slightly improve performance on the TinyFace database, while the tested fusion types were ineffective on the LFW database.