Search

The advanced technologies and sensors in smartphones has led to showcase their potential as a biometric sensor. In this work, we present the feasibility study and challenges in the path forward for using smartphone as a biometric sensor for iris recognition in visible spectrum. Especially, with a limited shelf-life of smartphones, it is anticipated to have enrolment and verification using different camera. In this work, we propose an improvement to segmentation scheme for contactless iris acquisition by approximating the radius range. The proposed method has resulted in a segmentation accuracy of 81%. We also propose various protocols for real-life verification scenarios using smartphones for visible spectrum iris recognition. Finally, results from an extensive set of experiments are presented to validate the anticipated challenges in using smartphone based iris recognition. Being the first of its kind, this work provides the benchmarking results for the smartphone iris database. The best EER is obtained for iPhone in indoor scenario with an impressive EER of 0.48%.

Multimodal biometric systems based on fingerprint and finger vein modality provide promising features useful for robust and reliable identity verification. In this paper, we present a robust imaging device that can capture both fingerprint and finger vein simultaneously. The presented low-cost sensor employs a single camera followed by both near infrared and visible light sources organized along with the physical structures to capture good quality finger vein and fingerprint samples. We further present a novel finger vein recognition algorithm that explores both the maximum curvature method and Spectral Minutiae Representation (SMR). Extensive experiments are carried out on our newly collected database that comprises of 1500 samples of fingerprint and finger vein from 150 unique fingers corresponding to 41 subjects. Our results demonstrate the efficacy of the proposed sensor with a lowest Equal Error Rate of 0.78%.

This paper proposes a new scheme for robust palmprint verification using sparse representation of Binarized Statistical Image Features (BSIF). Since palmprint comprises of rich set of features including principal lines, ridges and wrinkles, the use of appropriate texture descriptor is expected to accurately capture these information. To this extent, we explore the BSIF texture descriptor which codes each pixel of the given palmprint image in terms of binary strings based on the filter response. The BSIF learns the filter basis from the natural images by exploring statistical independence. We then use the Sparse Representation Classifier (SRC) on these BSIF features to perform the subject verification. Extensive experiments are carried out on three different large scale publically available palmprint databases. We then present an extensive analysis by comparing the proposed scheme with five different contemporary state-of-the-art schemes that reveals the outstanding performance.

The advancement in face recognition algorithm has a strong relationship with the availability of face databases that exhibit varying factors reflecting real life scenarios. The GUCLF face database is the first of its kind that can strongly influence the advancement in face recognition technology. In this paper, we introduce and describe our new face samples database collected using Lytro light field camera. The database consists of 200 reference samples and 303 probe samples collected from 25 subjects. The reference samples are collected in the controlled conditions using Canon EOS 550D DSLR camera. While probe samples are captured using both conventional digital camera (Sony DSC-S750) and Lytro light field camera. The probe samples are captured in three different scenarios: indoor, corridor and outdoor to include all possible real life conditions. In addition to the database description, this paper also elaborates on possible uses of the collected database and proposes a testing protocol. Further, we also present the quantitative results from the baseline experiments using the Kernel Discriminant Analysis (KDA).

Iris and Periocular biometrics has proved its effectiveness in accurately verifying the subject of interest. Recent improvements in visible spectrum Iris and Periocular verification have further boosted its application to unconstrained scenarios. However existing visible Iris verification systems suffer from low quality samples because of the limited depth-of-field exhibited by the conventional Iris capture systems. In this work, we propose a robust Iris and Periocular erification scheme in visible spectrum using Light Field Camera (LFC). Since the light field camera can provide multiple focus images in single capture, we are motivated to investigate its applicability for robust Iris and Periocular verification by exploring its all-in-focus property. Further, the use of all-in-focus property will extend the depth-of-focus and overcome the problem of focus that plays a predominant role in robust Iris and Periocular verification. We first collect a new Iris and Periocular biometric database using both light field and conventional camera by simulating real life scenarios. We then propose a new scheme for feature extraction and classification by exploring the combination of Local Binary Patterns (LBP) and Sparse Reconstruction Classifier (SRC). Extensive experiments are carried out on the newly collected database to bring out the merits and demerits on applicability of light field camera for Iris and Periocular verification. Finally, we also present the results on combining the information from Iris and Periocular biometrics using weighted sum rule.

In this paper, we propose a novel approach for reconstructing 3D face in real-life scenarios. Our main objective is to address the most challenging issue that involves reconstructing depth information from a video that is recorded from frontal camera of the smartphone. Such videos recorded using smart-phones impose lot of challenges, such as motion blur, non-frontal perspectives and low resolution. This limits the applicability of state-of-the-art algorithms, which are mostly based on landmark detection. This situation is addressed with the Scale-Invariant Feature Transformation (SIFT) followed by feature matching to generate consistent tracks. These tracks are further processed to generate a 3D point cloud using Point/Cluster based Multi-view stereo (PMVS/ CMVS). The usage of PMVS/CMVS will however fail to generate a dense 3D cloud points on the weak surfaces of face, such as cheeks, nose and forehead. This issue is addressed by multi-view reconstruction of these weakly supported surfaces using Visual-Hull. The effectiveness of our method is evaluated on a newly collected dataset, which simulates a realistic identification scenario using a smartphone.

On-the-fly person identification using face has been receiving more and more attention recently as it exhibits additional challenges that include uncontrolled poses, change in illumination and focus as compared to that of conventional face recognition. In this paper, we propose an efficient scheme that can be structured in two steps namely face detection and person identification to accurately address the challenges in “On-the-Fly” person identification. The proposed face detection scheme is based on the skin color region detection and a novel region based association scheme which is efficient and innovative in locating the evolving face in the video. We then, propose to carry out the person identification by employing Log-Gabor transform and Sparse Representation Classifier (SRC). Extensive experiments are carried out on the public available video surveillance `ChokePoint' dataset leading to the highest observed closed-set identification rate of 91% with efficient computational speed of 27 frames/minute on normal desktop.

Face recognition in a wide area surveillance scenario is a challenging problem as it requires multiple cameras, needs to process low resolution samples and high computation. In this paper, we experiment with a Light-Field Camera (LFC) to effectively address a wide area surveillance scenario. The interesting property of the LFC is that, it can provide different focus (or depth) images in a single capture that was not possible with any conventional camera (CCTV or PTZ) widely employed for surveillance applications. We further explore the property of multiple depth images rendered by LFC to perform the super-resolution reconstruction. To this extent, we first present a quantitative and comparative evaluation of well known super-resolution techniques. We then, present a new resolution enhancement scheme based on Discrete Wavelet Transform (DWT) as well as existing super-resolution scheme to capture the high frequency components from different depth images rendered by LFC and thereby improving the performance of the face recognition. Experiments are carried out on our new face dataset to bring out the merits and demerits of employing the light-field camera for face recognition in wide area surveillance applications.

Rising early in the morning has been a prescribed discipline of ancient Indian tradition. While there are no scientific studies comparing early rising volitionally versus circumstantially, selected studies on the latter (rising forcefully) have shown negative impact on an individual's peroformance. Hence the present study was undertaken to assess the influence of early rising (during Brahma-muhurtha) on tasks requiring attention and the ability to recall. Fifty four normal healthy male volunteers, with ages ranging from 16-22 years from a residential school were selected. They were randomly allocated to two groups (Brahma-muhurtha and control). They were assessed on day 1, day 10 and day 20 of the intervention, using a digit letter substitution task and verbal and spatial memory task. The Brahma-muhurtha group were asked to rise before 4:30 am in the morning based on the traditional Indian astrological calculations, while the control group were allowed to wake up just before 7 am which was their regular timing for waking. Brahma-muhurtha group after 20 days showed a significant improvement in the net scores for digit letter substitution task as well as scores for verbal and spatial memory tasks. The control group also showed an improvement in the memory task but not in the task requiring attentional processes. The present study suggests that rising early in the morning as described in ancient Indian tradition influences the process of attention and can improve the ability to recall.

This paper presents the new feature extraction scheme for accurate leaf classification. The proposed feature extraction scheme can be viewed as a combination of Gabor transform and Local Phase Quantization (LPQ) and we term this scheme as Local Gabor Phase Quantization (LGPQ). First, the Gabor magnitude images are obtained by convolving the given leaf image with Gabor filter with different scale and orientation. Then, we divide each of these Gabor magnitude images into number of sub-images of size $10 \times 10$. Then, we encode each of these sub-images using LPQ to capture the rich set of information and concatenated to form a single feature set. We then, use Principal Component Analysis (PCA) to reduce the dimension of the feature space. Finally, the reduced feature set is classified using Support Vector Machine (SVM). Extensive experiments are carried out on three different datasets with varying size and illumination to prove the efficacy of the proposed scheme.

Finger vein recognition has emerged as an accurate and reliable biometric modality that was deployed in various security applications. However, the use of finger vein recognition also indicated its vulnerability to presentation attacks (or direct attacks). In this work, we present a novel algorithm to identify the liveness of the finger vein characteristic that is presented to the sensor. The core idea of the proposed approach is to magnify the blood flow through the finger vein to measure its liveness. To this extent, we employ the Eulerian Video Magnification (EVM) approach to enhance the motion of the blood in the recorded finger vein video. Next, we further process the magnified video to extract the motion-based features using optical flow to identify the finger vein artefacts. Extensive experiments are carried out on a relatively large database that is comprised of normal presentations vein videos from 300 unique finger instances corresponding to 100 subjects. The finger vein artefact database is captured by printing 300 real (or normal) presentation image of the finger vein sample on a high-quality paper using two different kinds of printers namely laser and inkjet. Extensive comparative evaluation with four different well-established state-of-the-art schemes demonstrated the efficacy of the proposed scheme.

Face recognition systems assume that a person's face serves as the unique link to identify them. A morph attack happens when two people with similar facial features morph their faces together, resulting in a face image that can be identified as either of the two contributing individuals. Since the morphed image inherit enough visual traits from both individuals, both humans and automatic algorithms could be deceived by a morphed image. In terms of biometrics, changing one's appearance to impersonate a target identity is a direct attack on the security of face recognition systems. Defending against such attacks necessitates the ability to detect them as distinct identities from their target. Although they are not always visible in the image domain, many morphing algorithms introduce artifacts in the final image that can be used to detect morph attacks. Since various spectral images allow us to investigate low and high-frequency data separately, we can recognize and isolate these morphing abnormalities in the spatial frequency domain. For this research, we develop a new database that includes the morphed images created using three different techniques and spectral images in different spectral bands. This study studies the potential attack of various efficient face recognition systems from the newly created database using spectral images as a reference set. In addition, this thesis also investigates the human observer's ability to detect the morphed images while examining spectral images. Further, we evaluate the effectiveness of different MAD approaches using spectral bands imaging in order to detect differential morph attacks.

Mobai ønsket å utnytte sine biometriske tjenester til å bygge en OpenID Connect-leverandør med vedvarende lagring av biometriske data som både er sikkert og overholder prinsipper for personvern. Gjennom hele prosjektet ble Datatilsynets prinsipper for innebygd personvern og Microsoft Security Development Lifecycle benyttet for å gi et helhetlig personvernfokus gjennom alle utviklingstrinn. Systemet bruker en mikrotjeneste-arkitektur og bruker Kubernetes for å orkestrere containerne. For å sikre og beskytte brukernes biometriske data, ble en biometrisk malbeskyttelsesalgoritme basert på bloomfiltre implementert. Til slutt ble det foretatt en risikovurdering av systemet, som identifiserte aktuelle risikoer og potensielle tiltak for reduksjon av risikoen. Resultatet er en prototype som demonstrerer mulighetene av å bruke et autentiseringssystem som benytter biometri som autentiseringsfaktor, i tillegg til å oppfylle de strenge personvernforskriftene som er pålagt av General Data Protection Regulation (GDPR). Mobai wanted to leverage their biometric services to build a OpenID Connect Provider with persisted storage of biometric data which is both secure and adheres to privacy by design principles. Throughout the project, the Norwegian Data Protection Authority privacy by design guide and Microsoft Security Development Lifecycle were utilized to provide a holistic privacy focus throughout all development stages. The system uses a microservice architecture and utilizes Kubernetes for orchestrating the containers. In order to secure and protect the users biometric data, a biometric template protection algorithm based on bloom filters was implemented. Finally, a risk assessment of the system was conducted, which identified current risks and potential mitigation. The result is a prototype which demonstrates the opportunities of using an authentication system which leverages biometrics as the method of authentication, in addition to fulfilling the strict privacy regulations imposed by General Data Protection Regulation (GDPR).

Morphing attacks have posed a severe threat to Face Recognition System (FRS). Despite the number of advancements reported in recent works, we note serious open issues such as independent benchmarking, generalizability challenges and considerations to age, gender, ethnicity that are inadequately addressed. Morphing Attack Detection (MAD) algorithms often are prone to generalization challenges as they are database dependent. The existing databases, mostly of semi-public nature, lack in diversity in terms of ethnicity, various morphing process and post-processing pipelines. Further, they do not reflect a realistic operational scenario for Automated Border Control (ABC) and do not provide a basis to test MAD on unseen data, in order to benchmark the robustness of algorithms. In this work, we present a new sequestered dataset for facilitating the advancements of MAD where the algorithms can be tested on unseen data in an effort to better generalize. The newly constructed dataset consists of facial images from 150 subjects from various ethnicities, age-groups and both genders. In order to challenge the existing MAD algorithms, the morphed images are with careful subject pre-selection created from the contributing images, and further post-processed to remove morphing artifacts. The images are also printed and scanned to remove all digital cues and to simulate a realistic challenge for MAD algorithms. Further, we present a new online evaluation platform to test algorithms on sequestered data. With the platform we can benchmark the morph detection performance and study the generalization ability. This work also presents a detailed analysis on various subsets of sequestered data and outlines open challenges for future directions in MAD research., This paper is a pre-print. The article is accepted for publication in IEEE Transactions on Information Forensics and Security (TIFS)

Biometric authentication has been a vital method of human identification across a wide range of applications. It is a quick and secure authorization process in applications such as border control and banking transactions. Recent advances in technology have triggered the use of embedded biometrics in smartphones and handheld devices. For reliable authentication, biometrics perform better when compared to traditional techniques such as passwords. Moreover, its advantages, such as permanence and uniqueness, have increased the growth of biometrics in everyday usage. However, there are certain limitations to biometric systems in providing optimal performance. These limitations play a crucial role in formulating artefacts to conceal or recreate the identity of individuals. Therefore, this raises concerns about the robustness of a biometric system and questions the accuracy of biometric recognition. In a real-world scenario, an end-user biometric system is protected from tampering by external sources. However, the only source of interaction is through the data capturing sensor. Thus, external artefacts, namely presentation attacks, have been a severe threat to biometric systems. On the other hand, the internal dependencies typically come from the limitations of the hardware and software deployed in the biometric verification workflow. Examples of internal dependencies vary from noise in biometric data to dependencies of biometric algorithms. In this thesis, we have focused on improving the generalization of biometrics by working on some of the problems caused by presentation attacks and internal dependencies in biometrics. The key challenges in audio-visual biometrics were identified, and research objectives were designed for this thesis. The vulnerabilities in audio-visual biometrics are observed with the help of a thorough review of existing recognition and presentation attack detection methods. An exhaustive and comprehensive study along with a comparison and categorization of state-of-the-art methods have resulted in a novel dataset. The dataset includes different attributes, which provide the scope to perform extensive experiments to understand dependencies and vulnerabilities. The thesis proposes a fusion of texture features based iris presentation attack detection algorithm, with results showing superior performance. Further, the cross-dataset experiments led to an empirical evaluation of vulnerabilities in iris biometrics due to presentation attacks. In speaker recognition research, voice impersonation, language dependency and audio replay attacks cause high vulnerability. This thesis proposes a novel voice impersonation dataset with three different languages. The impact of voice impersonation as an attack is evaluated on the state-of-the-art speaker verification methods. Further, the speaker’s language mismatch in the enrollment and testing steps of speaker verification is examined. The recent progress in smartphone usage is also reflected in high-quality speakers and microphones. A replay attack dataset is created with multiple smartphones as playback and recording devices, and vulnerability is examined. The thesis examines the generalizability of biometric algorithms to improve the robustness of biometric recognition. The results from the proposed methods are evaluated with extensive experiments and detailed examinations of both publicly available databases and new datasets created in this work. In conclusion, the thesis proposes novel methods and approaches to examining the vulnerabilities in audio-visual biometrics, presentation attack detection (PAD) in iris and voice biometrics, the study of language dependency and audio replay attacks. The methods presented are valuable contributions to the research fields in developing robust smartphone biometric methods by addressing vulnerabilities from multiple sources.

Following the introduction of image manipulation tools such as Adobe Photoshop in the early 2000s, the public trust in image authenticity dropped and the need for the development and deployment of image authentication techniques became apparent. Recently, we face a similar situation for video content as photo-realistic video manipulation tools like Deepfake are becoming available and within the reach of the general public as well as bad actors. In human to human communication, face and voice modalities play a crucial role, and not surprisingly, the same modalities are most under attack by forgers. Historically, the task of audiovisual content authentication was the focus of the field of multimedia forensics, with more than 15 years of accumulated literature. Following the increase in the popularity of biometric systems in practice, these systems have also faced similar challenges and felt the need for content authentication. Consequently, the field of presentation attack detection is born to protect biometric systems against fake biometric presentations. Due to the parallel nature of the presentation attack detection problem, defined as protecting a biometric system from presentation attacks, to the audiovisual content authentication problem, defined as protecting the viewer from fake content, the field of biometric presentation attack detection can provide a solid basis for approaching the multimedia authentication problem. The primary objective of this thesis is to address the audiovisual content authentication problem on the face modality by vulnerability assessment and mitigation of detected vulnerabilities with reliance on biometric and presentation attack detection knowledge. To this end, after producing a taxonomy of existing generation techniques, subjective tests are done to assess the vulnerability of viewers to the most prevalent generation techniques with reliance on data collected from the wild. Following this process, the generation techniques the viewers are most susceptible to were identified. The discovered vulnerabilities are then mitigated individually by the introduction of effective detection techniques that outperform existing solutions. Furthermore, the vulnerability of existing general-purpose detection methods was analyzed and it was discovered that these methods show limited generalization capacity when faced with new generation methods. To mitigate this vulnerability, with reliance on an anomaly extraction approach, a generalizable detection method is introduced and empirically evaluated against the state-of-theart methods. Additionally, all the datasets that are collected during the course of this thesis work are made publicly available to stimulate further research on this topic.

Sosiale medier har blitt svært populære i den siste tiden. Blant alle fordelene som sosiale medier bragt med seg, er det også noen problemer. Ett av de største problemene som har kommet med den økte bruken av sosiale medier er nettmobbing. Ny statistikk viser at nettmobbing er et problem som øker i omfang år etter år. Problemet er at sosiale medier skaper så mye data at det er svært vanskelig å fange opp og stoppe nettmobbing ved hjelp av manuell overvåkning. Dette har motivert forskere til å forsøke å finne en automatisert metode for å fange opp tilfeller av nettmobbing. I de siste årene så har det blitt populært blant forskere å bruke en type kunstig intelligens nettverk kalt Long-Short-Term-Memory Neural Network til dette formålet. I denne masteroppgaven vil vi presentere vår empiriske evaluering av forskjellige typer kunstig intelligens nettverk, for å undersøke hvordan man bør designe disse nettverkene for å maksimere deres evne til å fange opp nettmobbing. Vi har foretatt to omfattende eksperimenter med Long-Short-Term-Memory Neural Networks, hvor vi har testet 56 forskjellige design, for å se hvilke som fungerer best. Disse eksperimentene har blitt gjennomført som faktoriale eksperiment, som betyr at et sett av variabler har blitt endret i samspill med hverandre. Dette har gjort det mulig for oss å studere årsak-virkning sammenhenger mellom variablene på en mye bedre måte enn hva som er mulig ved å kun studere tidligere publisert forskning. For å gjøre eksperimentene i denne studien så relevant som mulig for fremtidig forskning innen feltet, så har vi basert hele vår Natural Language Processing pipeline som leder opp til kunstig intelligens nettverkene på eksisterende «state of the art» forskning innen feltet. Vi går også igjennom vår Natural Language Processing pipeline i detalj, slik at den lett kan bli gjenbygd av utenforstående, og dermed brukt til fremtidig forskning på andre variabler enn de vi tester. I våre eksperimenter så tester vi fire hoved variabler; komposisjon av forskjellige typer nettverkslag, stabling av like typer nettverkslag oppå hverandre, hvor mange generasjoner nettverket trenes og aktiveringsmekanisme for de to forskjellige klassene. Våre funn viser at den beste komposisjonen av nettverkslag er å kombinere lagene fra et typisk Convolutional Neural Network med Long-Short-Term-Memory nettverkslag. Vi fant også ut at uansett hvilken komposisjon av nettverkslag man bruker så vil man forbedre ytelsen til det kunstige nettverket ved å stable to like sett med lag oppå hverandre. Som forventet, så fant vi også ut at et nettverk som er trent tilstrekkelig yter bedre enn et nettverk som ikke er trent nok, men det er ikke noen sammenheng mellom hvor mye trening et nettverk krever og hvor godt det yter. Til slutt så beviste vi også at det ikke er noen grunn til å alltid velge Softmax aktivering i et kunstig nettverk, slik som de fleste forskerne har en tendens til å gjøre. Dette fordi vår studie viser at et nettverk med en aktiveringsmekanisme som etterligner hvordan Support-Vector-Machines modeller virker, yter bedre enn Softmax aktivering. Social media usage has skyrocketed the last decades. Along with all the upsides of social media, there are also a few downsides. One of the most problematic downsides has been the rise of cyberbullying. New statistics shows that this problem has kept on escalating in the recent years. The problem with detecting and preventing cyberbullying is that social media generates so much data that it is virtually impossible to detect cyberbullying by performing manual inspections. This has motivated researchers to come up with new automated methods of detecting cyberbullying. In the most recent years, it has become popular among researchers to utilize a type of artificial Neural Network called Long-Short-Term-Memory Neural Network. In this thesis we present the empirical evaluation of different types of artificial Neural Network to see how the network can be designed to maximize its performance with regards to detecting cyberbullying. To find the best design of a Long-Short-Term-Memory Neural Network we conduct two experiments, testing 56 different designs. These experiments have been designed as factorial experiments, meaning that a set of variables has been modified in combination with each other. This lets us track any cause and effect relationships between the different variables in a much more accurate way, than what is possible by studying the existing research within the field. In order to make the results of this study as relevant as possible to the state-of-the-art cyberbullying research, we base the whole Natural Language Processing pipeline that leads up to the artificial Neural Network on the existing state-of-the-art research. We also explain this Natural Language Processing pipeline in depth, such that it can be recreated, and used for continuous study of other variables that are not tested by us, in future research projects. Four main variables are tested in this thesis; composition of different layer types, stacking of equal layer types, training epochs and activation mechanisms for the classes. Our findings show that the best composition of layer types are found by combining a traditional Convolutional Neural Network with a Long-Short-Term-Memory Neural Network. We also find that all of the Neural Networks, no matter what layer composition, benefit from going from no stacking to stacking its layers twice. As expected, we also find that Neural Networks perform better when properly trained, but there is no correlation between how much training a model needs and how well it performs. Lastly, we prove that there is no reason to blindly choose Softmax as the activation for the classes in the Neural Network models, as most of the researchers tend to do. Our findings show that using an activation mechanism that mimics the mechanisms of a Support-Vector-Machines classifier outperforms the Softmax activation, with all four different layer compositions of the Neural Network that we tested.