An Educational Innovation Project Focused on the Implementation of Biometrics in Portable Devices

Depósito de investigación de la Universidad de Sevilla https://idus.us.es/ This is an Accepted Manuscript of an article published by IEEE in XVI Congreso de Tecnología, Aprendizaje y Enseñanza de la Electrónica (TAEE) on 01 August 2024, available at: https://doi.org/10.1109/TAEE59541.2024.10604972 “© 2021 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other Works” An educational innovation project focused on the implementation of biometrics in portable devices Rosario Arjona, Paula López-González, Javier Arcenegui, and Iluminada Baturone Instituto de Microelectrónica de Sevilla (IMSE-CNM) Universidad de Sevilla, Consejo Superior de Investigaciones Científicas (CSIC) Seville, Spain {arjona, paula, arcenegui, lumi}@imse-cnm.csic.es Abstract—This work describes an educational project that explains how to extract biometric features from biometric samples acquired with microelectronic sensors suitable for portable devices (wearables and mobile phones) by satisfying low cost and reduced size. Specifically, fingerprints, ECG (electrocardiogram) and pulse signals, faces and veins are considered as biometric traits. Acquisition methods associated with this are fingerprint sensors, ECG electrodes and pulse sensors for wearables (in this work, a Raspberry Pi), and cameras for the acquisition of faces and veins in mobile phones. Then, it is explained how features are extracted and compared using biometric recognition algorithms selected to be implemented in portable devices: the feature QFingerMap16 (QFM16) for fingerprints, ECG and pulse waves, FaceNet embeddings for faces, and SIFT (Scale Invariant Feature Transform) for veins. Finally, it is explained how the features extracted are extensively evaluated by using public databases that contain samples acquired with sensors suitable for portable devices. Theoretical and experimental material for this research line and application field are distributed in four sessions lasting 4 hours and 30 minutes. Keywords— Biometrics, Portable Devices, Microelectronic Sensors, Educational Innovation I. INTRODUCTION This work was developed in the context of the educational innovation project “Research Students” granted by the Consejería de Desarrollo Educativo y Formación Profesional of the Andalusian Government. The target were 10 students in the fourth year of Secondary School and in the first year of High School with specialization in Science and Technology from 5 high schools. In addition, the realizations were also considered for the 36 students in the Technology specialization of the Master's Degree in Secondary School Teaching, Vocational and Language Teaching. The contents of this work are related to subjects where the topic “Electronic Technology” is taught. Over the past several years, there has been a significant increase in the use of portable devices, such as wearables and mobile phones, which have become essentials in our daily lives [1]. The rise of this trend is driven by technological progress, together with the growing need for mobility and convenience. In addition to the increasing number of portable devices, there is a growing acknowledgment of the significance of biometric security mechanisms. These mechanisms provide improved verification and authorization capabilities compared to traditional techniques like passwords or PINs. In an era of growing digitalization, it is crucial to educate individuals about biometric security and improve their comprehension of the importance of safeguarding sensitive data and preserving privacy. Biometrics is a discipline that includes many methods used to recognize individuals by their distinct physical, physiological or behavioral traits [2]. The biometric process normally has multiple steps, including acquisition, preprocessing, feature extraction, and matching. To carry out biometric recognition, two essential procedures must be undertaken: enrollment and verification. During the enrollment process, individuals acquire a biometric sample, preprocess it if needed, and extract the biometric features. These features are frequently stored in a database. During the verification step, users undergo the same process as in the enrollment step. However, instead of storing the extracted data, they are compared with the stored ones to obtain a score that determines the final result. Additionally, it is explored the selection of acquisition methods that are appropriate for portable devices. Wearable devices, such as smartwatches or health bracelets, depend on small sensors because of limited space availability. For example, capacitive sensors are used to collect fingerprints, while electrodes and pulse sensors are employed to obtain physiological signals such as electrocardiogram (EEG) and pulse, respectively. Our work also includes exploring the use of cameras to collect facial and vein biometric data on mobile phones. Furthermore, the evaluation of biometric samples is crucial for determining the efficiency and reliability of biometric systems. Hence, suitable algorithms should be selected for extracting and matching features, specifically designed for portable devices. In this work, different algorithms for each biometric trait are studied. Moreover, the careful choice of assessment criteria is essential for accurately evaluating the effectiveness and efficiency of these algorithms in real-life situations. In order to achieve this aim, comparisons are classified into two categories: genuine and impostor. Genuine comparisons involve samples obtained from the same person, while impostor comparisons involve samples obtained from different individuals. Following the FVC (Fingerprint Verification Competition) protocol [3], for a specific biometric database, genuine comparisons are performed with all the samples of the same person, and impostor comparisons are performed with the first sample of different people. The False Non-Match Rate (FNMR) and False Match Rate (FMR) are often used metrics to assess the performance of recognition systems. FNMR is also referred to as the False Fig. 1 Fingerprint sensor FPC1011F3 connected to a Raspberry Pi. Rejection Rate (FRR), while FMR is known as the False Acceptance Rate (FAR). The FNMR measures the number of genuine comparisons that are not matched. On the other hand, the FMR measures the number of impostor comparisons that are matched. The equal error rate (EER) is the point at which the FNMR and the FMR cross, offering valuable information about the performance of the system. This work seeks to offer insights on the importance and applicability of biometric security in the realm of portable devices. The rest of the sections are structured as follows. Section II describes different acquisition methods for certain biometric traits used in this work, such as fingerprints, ECG and pulse, faces, and veins. Section III outlines the selected algorithms for each biometric trait, specifically designed for portable devices with less complexity while still ensuring strong recognition performance. Section IV provides results using public biometric databases, specifically discussing the results of their implementation and difficulties encountered. These details are further elaborated in Section V. Finally, Section VI gives conclusions. II. ACQUISITION METHODS FOR PORTABLE DEVICES With this section, the students are introduced to acquisition methods suitable for wearable and mobile devices according to the biometric traits selected. A. Acquisition Methods of Fingerprints There are several types of fingerprint sensors: optical, capacitive, thermal and ultrasound sensors. The selection more suitable for portable devices should consider size, recognition accuracy and cost. The most extended fingerprint sensors for portable devices are capacitive sensors. In this work, we employ the Fingerprints FPC1011F3 capacitive sensor [4], which satisfies the requirements for portable devices. The fingerprint image size is 200x152 pixels with grayscale values of 8 bits. Datasheet is available so that the students can obtain information about the sensor specifications. The pins of the sensor (data out, data in, chip select, clock, reset, voltage supply, ground, ESD drain) are connected to the pinout of a Raspberry Pi as illustrated in Fig. 1. The communication is established by SPI (Serial Peripheral Interface). Functions implemented in C++ for configuration of the SPI (bit rate to 4000000) and for reading the pixels of each row of the fingerprint image (taking into account that 8 pixels are captured simultaneously) are provided to the students. The fingerprint image is saved in BMP format. B. Acquisition Methods of ECG and Pulse Signals ECG signals are captured through electrodes of different materials (mainly, gelled Ag/AgCl or stainless steel) which acquire the electrical activity of the heart. Pulse signals are captured through photoplethysmography sensors which include a LED to emit a NIR (Near-InfraRed) light (red or green of ~550nm) and a photodetector to measure changes in blood flows by means of the transmission or reflection of the light. Oxygenated hemoglobin, which is present in the blood of the arteries, absorbs NIR, while if there is no hemoglobin, the light is reflected. In this work, the platform BITalino is employed to acquire ECG and pulse signals by using the modules for ECG and pulse sensors. BITalino is an affordable and open-source biosignals platform designed for education and prototyping [5]. 3-lead or 2-lead accessories allow connecting electrodes. Datasheets are available so that the students can obtain information about the specifications of the electrodes and BITalino platform. The pins of the BITalino platform (RX, TX and CTS) are connected to the pinout of a Raspberry Pi as illustrated in Fig. 2. The communication is established by UART (Universal Asynchronous Receiver / Transmitter). C++ codes are provided to the students, which include functions for the BITalino connection, to establish the acquisition frequency to 1000 Hz, and to read the data through two channels (for ECG and pulse signals). When the acquisition is stopped by pressing a key, data are saved in a text file. C. Acquisition Methods of Faces In order to obtain faces, we utilize two smartphones, specifically the Huawei P40 and Xiaomi Redmi Note 8. These smartphones are equipped with advanced camera systems that have the ability to capture high-resolution photographs. Using the front-facing cameras of these smartphones, we employed a face acquisition procedure specifically created to collect facial photos in different lighting conditions. The Huawei P40 is well-known for its advanced camera technology, which includes features like high dynamic range (HDR) imaging and AI-powered scene recognition. The frontal camera has 32 MP and Time-Of-Flight (ToF) technology. Similarly, the Xiaomi Redmi Note 8 features a proficient camera configuration, comprising a highresolution sensor and software improvements specifically designed for portrait shooting. In this smartphone, the frontal camera has 13 MP, thus students can compare both cameras and their reliability. D. Acquisition Methods of Veins Conventional approaches for acquiring vein patterns usually include specialized NIR illumination in the range of 750-1000 nm and cameras. It is essential to understand that NIR enables the visibility of veins because hemoglobin present in the veins absorbs NIR, while the other tissues in the body let it pass through. Depending on where the Fig. 2 BITalino platform for the acquisition of ECG and pulse signals connected to a Raspberry Pi. ECG module is connected to a 2-lead accessory with gelled Ag/AgCl electrodes. Pulse module is connected to a photoplethysmography sensor. Fig. 3 Wrist vein pattern acquisition using the Xiaomi Redmi Note 8 smartphone. illumination and cameras are located, there are two possible acquisition mechanisms: reflection or transmission. Nevertheless, our strategy [6] diverges from the conventional approaches as we utilize the inherent functionalities of widely available smartphones, namely the Huawei P40 and Xiaomi Redmi Note 8, without making any alterations or using dedicated hardware, i.e., using their rear cameras of 32 MP and 48MP, respectively. We employ a commercial Android application called IRVeinViewer. This application employs the rear cameras of smartphones and performs image processing techniques to imitate near-NIR images and improve the viewing of veins. This program also triggers the flashlight when required, hence improving image quality in specific lighting settings. Due to the limitations of this procedure, only the most prominent and superficial veins near the skin can be accessed, such as hand dorsals and wrists. We have created a supplementary application that connects with IRVeinViewer, making the process of obtaining data more efficient for the students. This program retrieves the processed photos from IRVeinViewer and stores them in the internal memory of the device. It superimposes a rectangular shape on the screen of the smartphone, helping students to accurately locate the body part and ensuring that the image capture is consistent. Our program enhances the student experience by simplifying the acquisition process with real-time feedback and automated selection of the Region Of Interest (ROI). Fig. 3 shows the acquisition of wrist vein patterns following this approach. We conducted our data gathering method by capturing photographs in several settings, such as artificial light, darkness, and natural light, in order to evaluate the strength and adaptability of the approach. III. PROCESSING ALGORITHMS OF BIOMETRIC SAMPLES Once biometric samples are acquired, to illustrate the following steps of the biometric recognition process, the students are introduced to feature extraction and comparison algorithms. These algorithms are selected to be suitable for wearable and mobile devices. The algorithm codes are provided to the students so that they can study the implementations, perform parameter modifications and use them with the acquired samples. A. Extraction and Matching of Fingerprint Features In this work, we employ a fingerprint feature named QFingerMap16 (QFM16) which is very suitable for constrained-resource wearable devices, as proven in [7] through an implementation based on C++. The QFM16 feature is based on a subsampling of a directional image window centered at the convex core (a singular point that is included in all fingerprint classes). The directional image is extracted by determining a direction value (among 16 possible values) after applying a 9x9-pixel neighborhood operator for each pixel. Isolated and noisy directions are removed by smoothing the direction values with a 27x27-pixel maximum operator. The detection of the convex core is based on the intersection of homogeneous direction regions. A 128x128pixel window is centered at the convex core and a down sampling by a factor of 8 is applied. Fig. 4 shows the steps of the extraction process. Since the directional image contains 16 direction values (which can be coded with 4 bits) and the window extracted contains 16x16 pixels, a QFM16 is composed of 1024 bits. The comparison of QFM16s is based on the fusion distance (𝐹𝑑𝑖𝑠𝑡) based on the minimum operator of Hamming distances (𝐻𝑑𝑖𝑠𝑡) obtained from all the possible combinations of template and query QFM16 vectors: 𝐹𝑑𝑖𝑠𝑡 =1 1024∙min 𝑒𝑣=1,…,𝑗 𝑞𝑣=1,…,𝑘[𝐻𝑑𝑖𝑠𝑡(𝑄𝐹𝑀16𝑒𝑣,𝑄𝐹𝑀16𝑞𝑣)] () with: 𝐻𝑑𝑖𝑠𝑡 =∑𝑄𝐹𝑀16𝑖 𝑒⨁𝑄𝐹𝑀16𝑖 𝑞 𝑛 𝑖=1 () B. Extraction and Matching of ECG and Pulse Features Since ECG and pulse waves can be considered directly as features, in this work we consider them in order to simplify the feature extraction process. Fig. 5 shows examples of ECG and pulse waves by including the main wave peaks. For wearable devices, we select algorithms employed in [8] which Fig. 4 Extraction process of the fingerprint feature QFM16. 16-value directional and convex core detected 128x128 distinctive window 16x16 downsampled feature 510 15 20 25 30 5 10 15 20 25 30 QFM16 fingerprint image (a) (b) Fig. 5 Examples of ECG (a) and pulse (b) waves with R and systolic peak depicted, respectively. R peak R peak Systolic peak Systolic peak Fig. 6 Facial embedding extraction process. Face acquisition Facial detection and cropping Embeddings extraction [0.234, 0.127, -0.023 …. 0.263, -0.145] Embeddings binarization 1110010110… 111000 were implemented in Python. Steps for the extraction of ECG and pulse waves are similar. Firstly, a normalization should be performed. Secondly, denoising and peak detection should be applied. For ECG: two median filters with sliding windows of 200 ms and 600 ms, a low-pass Finite Impulse Response of order 300 with a cutoff of 15 Hz, and the detection of R peaks. For pulse: a band-pass Butterworth filter of order 4 with cutoff frequencies of 1 and 8 Hz, and the detection of systolic peaks. R and systolic peaks are detected by the location of maximum values from wave onsets. Finally, segmentation of fixed-time windows of 600 ms is performed from which 360 samples are extracted and three fixed-time windows are averaged to obtain waves without noise. Comparison of enrollment and query ECG and pulse waves is performed by the normalized cosine distance (𝐶𝑑𝑖𝑠𝑡): 𝐶𝑑𝑖𝑠𝑡 =1− 𝑊𝑒∙𝑊𝑞 ‖𝑊𝑒‖2‖𝑊𝑞‖2 2=1− ∑𝑊𝑖 𝑒∙𝑊𝑖 𝑞 𝑛 𝑖=1 √∑𝑊𝑖 𝑒2 𝑛 𝑖=1 √∑𝑊𝑖 𝑞2 𝑛 𝑖=1 2 () C. Extraction and Matching of Face Features Most extended face features are based on Convolutional Neural Networks (CNNs) since the recognition accuracy results are much better than the results obtained with other types of face features. The load and inference of CNNs for face biometric recognition is supported by, for example, the library TensorFlowLite, which is suitable for implementations on mobile devices [9]. After acquiring the face image, a 160x160-pixel face is detected and cropped with a pre-trained BlazeFace CNN. Then, 128 floating-point embeddings are extracted with a pre-trained FaceNet CNN. Fig. 6 shows the extraction process of embeddings. The floating-point embeddings can be binarized with 3 bit-codes (000, 001, 011 and 111) by applying Linearly Separable Subcode (LSSC). In that case, each feature vector contains 384 bits. The comparison of enrollment and query floating-point embeddings is performed by using the Euclidean distance (𝐸𝑑𝑖𝑠𝑡): 𝐸𝑑𝑖𝑠𝑡 =∑(𝑒𝑖𝑒−𝑒𝑖𝑞)2 𝑛 𝑖=1 () The comparison of binary embeddings is performed by using the Hamming distance as in (2). D. Extraction and Matching of Veins Features Among the features extracted from veins, we select SIFT (Scale-Invariant Feature Transform) features whose implementation in C++ was proven for mobile devices in [6]. SIFT features extract local maximum points (keypoints) through the space and frequency domains. Each keypoint is described by horizontal and vertical locations, scales, and orientations. The steps of the extraction process are: (1) image rescaling to 300x300, (2) ROI (Region Of Interest or foreground) extraction, (3) image enhancement (CLAHE (Contrast Adaptive Limited Histogram Equalization) twice and subsequently Gaussian, median, and average filters with a kernel size of 11x11), and (4) SIFT keypoints extraction. The matching process is facilitated by the Fast Library for Approximate Nearest Neighbors (FLANN), a flexible tool for quickly finding approximate nearest neighbors in highdimensional spaces. FLANN calculates the normalized number of matches between keypoints, which serves as a reliable indicator of similarity between images. This method not only provides accurate matching, but also ensures computational efficiency, making it suitable for real-time applications. To avoid false matches, a verification process is implemented to ensure that the connecting segments between matching pairs have similar lengths and orientations. Fig. 7 shows keypoints detected in two veins images and the comparison. IV. EVALUATION USING PUBLIC DATABASES Since the biometric samples acquired in Section II should be treated confidentially and privately, should be deleted from the memories of the devices, and should not be distributed or published, public databases are employed to illustrate the evaluation of the recognition algorithms. In addition, public databases allow for extensive evaluation of the algorithms. The selection of public databases tried to include samples acquired with sensors suitable for portable devices if the databases were available. Evaluation codes and distributions of database samples are provided to the students so that they can obtain recognition results. For fingerprint samples, the database FVC 2000 DB2a was selected [10]. FVC 2000 DB2a is a database created for the Fingerprint Verification Competition in 2000 by employing the low-cost capacitive sensor ST Microelectronics TouchChip (similar to the fingerprint sensor considered in Section II). DB2a contains 100 fingers and 8 impressions per finger from 20 to 30 year-old students (about 50% male). Image size is 256x364. The images were taken from untrained people in two different sessions and no efforts were made to assure a minimum acquisition quality. The acquired fingerprints were manually analyzed to assure that the maximum rotation is approximately in the range [-15°, 15°] and that each pair of impressions of the same finger has a nonnull overlapping area. For ECG samples, the database CYBHi was considered [11]. ECG signals were acquired from 63 individuals (14 males and 49 females, with an average age of 20.68 ± 2.83 years) without health problems. Ag/AgCl electrodes were located at two fingers (one from the left and another from the right hand). The acquisition was performed at a sampling rate (a) (b) (c) Fig. 7 SIFT extraction of two samples of the same individual (a) and b), and the comparison between them (c). of 1000 Hz for 2 minutes in two sessions separated during 3 months. For this work, the second session was considered with a duration required to extract 8 waves. For pulse samples, the database HiMotion was considered [12]. Pulse signals were acquired from 27 individuals (18 males and 9 females, with average ages of 23.4 ± 2.5 years) without health problems. The pulse sensor was placed in the distal phalange of the ring finger of the non-dominant hand. The acquisition was performed at a sampling of 256 Hz during 45 minutes while different tests were applied (association, concentration, discovery, intelligence and memory). For this work, the signals collected during the memory test were considered. For facial samples, we utilized the Labelled Faces in the Wild (LFW) database [13]. This database contains facial samples collected from the Internet, consisting of 5,749 individuals and a total of over 13,000 samples. LFW is widely recognized as a leading database for testing facial recognition algorithms. It presents significant challenges due to the intrinsic variations in poses, angles, lighting conditions, and aging present in the collected images. This diversity present in LFW accurately reflects real-world scenarios, making it an invaluable resource for evaluating the robustness and generalization capabilities of facial recognition systems. For vein samples, we employed the VERA-PalmVein database [14]. This database is purposefully designed for doing research on palm vein recognition and acts as a complete benchmark for assessing the effectiveness of palm vein recognition algorithms. The VERA-PalmVein dataset comprises a wide range of palm vein images obtained from 110 individuals, with samples captured from both hands. The images have a standardized size of 480x680 pixels. The database was gathered in two sessions, with each session capturing five photos of each hand. The database comprises 40 females and 70 males, all aged between 18 and 60, with an average age of 33. The palm vein images were captured using a contactless prototype sensor for the palm vein. This sensor was composed of the following components: an ImagingSource camera, a Sony ICX618 sensor, LEDs that emit infrared light at a wavelength of 940 nm, a HC-SR04 ultrasound sensor to measure the distance between the user's hand and the camera lens, and a LED signal that guides the user to the appropiate hand position for the acquisition. In this study, the entire database consisting of 2,200 photos was considered. Recognition results after the evaluation with the algorithms described in Section III and the public databases mentioned above are included in Table I. Recognition results show that fingerprint and face features are more accurate than ECG, pulse and veins features. However, fingerprint and face are more vulnerable to spoofing attacks and other biometric traits such as ECG, pulse and veins could overcome this drawback. V. RESULTS OBTAINED AND PROBLEMS ENCOUNTERED Four sessions, each lasting 4 hours and 30 minutes, were scheduled and took place at the Instituto de Microelectrónica de Sevilla (IMSE). First session activities were a talk about security and biometric recognition, and the search of information related to the topic on the Internet. Second session activities were a talk about microelectronic sensors focused on biometrics, covering topics such as acquisition methods, and the use of wearable devices and mobile phones to acquire biometric samples. Third session activities were a talk about biometric recognition algorithms, and the evaluation of these algorithms using public databases. Fourth session activities were the generation of project final documents: a paper, a poster and a presentation to be presented in an educational conference (Jornada Científica en Dos Hermanas). A relevant issue is that, due to the personal and sensitive nature of biometric data, biometric samples acquired in the third session were employed in a confidential manner, without being stored or publicly shared, and without being a risk for the health. The students provided explicit, informed consent to participate in these sessions after receiving clear and understandable information about the project. The students showed interest in the topics covered and the sessions were very participatory. The main problem encountered was the duration of the project. More sessions were required to teach the contents in more depth. VI. CONCLUSIONS AND FUTURE WORK The project was successfully carried out, resulting in significant knowledge gained in the field of biometric security in relation to portable devices. We showed that it is possible and effective to use biometric recognition on wearable and mobile devices by studying acquisition methods, selecting algorithms, and evaluating metrics. The project not only improved the understanding of biometric security, but also offered essential practical experience for the participating students, cultivating their abilities in research, data analysis, and problem-solving in the realm of technology and biometrics. In future set ups, we could investigate combining various biometric features to improve reliability and security. Through the integration of various biometric modalities, such as fingerprints, facial features, and vein patterns, we can develop strong multi-modal recognition systems that can overcome the constraints associated with individual biometric traits. This method shows potential for attaining greater levels of accuracy and resistance against spoofing attacks, hence strengthening the overall security of portable devices and other applications. ACKNOWLEDGMENT This research was conducted thanks to Grants CPP2022009796 and PDC2023-145873-I00 funded by MICIU/AEI/10.13039/501100011033 and the “European Union NextGenerationEU/PRTR”, and Grant PID2020TABLE I. RECOGNITION ACCURACY OF DIFFERENT BIOMETRIC FEATURES EVALUATED WITH PUBLIC DATABASES Biometric Feature Database EER (%) Fingerprint QFM16 FVC 2000 DB2a 2.76 ECG Wave CYBHi 9.86 Pulse Wave HiMotion 13.84 FaceNet FloatingPoint Embeddings LFW 0.83 FaceNet Binary Embeddings LFW 1.18 Veins SIFT VERA-PalmVein 9.43 119397RB-I00 funded by MICIU/AEI/10.13039/501100011033. REFERENCES [1] K. H. Kwon, K. R. Park. “Wearable and implantable sensors for biomedical applications”, Annual Review of Analytical Chemistry, vol. 9, pp. 115-135, 2016. [2] A. K. Jain,, A. Ross, “Handbook of biometrics”, Springer Science & Business Media, 2008. [3] FVC (Fingerprint Verification Competition) Ongoing: https://biolab.csr.unibo.it/fvcongoing/UI/Form/Home.aspx [4] Fingerprints FPC1011F3 capacitive sensor: https://www.fingerprints.com/technology/hardware/sensors/fpc1011f3 / [5] BITalino platform: https://www.pluxbiosignals.com/collections/bitalino [6] P. López-González, I. Baturone, M. Hinojosa, and R. Arjona, “Evaluation of a Vein Biometric Recognition System on an Ordinary Smartphone”, Appl. Sci., vol. 12, 3522, 2022. [7] R. Arjona, J. Arcenegui and I. Baturone, “How to Implement a Fingerprint Recognition Algorithm into a Wearable Device”, XIV Technologies Applied to Electronics Teaching Conference (TAEE), Porto, Portugal, pp. 1-7, 2020. [8] M. Martinho, A. Fred and H. Silva, “Towards Continuous User Recognition by Exploring Physiological Multimodality: An Electrocardiogram (ECG) and Blood Volume Pulse (BVP) Approach”, International Symposium in Sensing and Instrumentation in IoT Era (ISSI), Shanghai, China, pp. 1-6, 2018. [9] P. López-González, R. Arjona, R. Román, and I. Baturone. “A facial authentication system using post-quantum-secure data generated on mobile devices”, 28th Annual International Conference on Mobile Computing And Networking (MobiCom), Association for Computing Machinery, New York, NY, USA, pp. 803–805, 2022. [10] D. Maio, D. Maltoni, R. Cappelli, J. L. Wayman, and A. K. Jain, “FVC2000: Fingerprint Verification Competition”, IEEE Transactions on Pattern Analysis Machine Intelligence, vol.24, no.3, pp.402-412, 2002. [11] H. Plácido da Silva, A. Lourenço, A. Fred, N. Raposo, and M. Airesde-Sousa, “Check Your Biosignals Here: A new dataset for off-theperson ECG biometrics”, Computer Methods and Programs in Biomedicine, vol. 113, no. 2, pp. 503-514, 2014. [12] H. Gamboa, H. Plácido da Silva, and A. Fred, “HiMotion: a new research resource for the study of behavior, cognition, and emotion”, Multimedia Tools and Applications, vol. 73, 2014. [13] G. B. Huang, M. Ramesh, T. Berg, and E. Learned-Miller, “Labeled Faces in the Wild: A Database for Studying Face Recognition in Unconstrained Environments”, University of Massachusetts Technical Report 07-49, 2007. [14] P. Tome, and S. Marcel, “On the Vulnerability of Palm Vein Recognition to Spoofing Attacks”, IAPR International Conference on Biometrics (ICB), 2015.