Abstract: Online handwritten signature verification is an important behavioral biometric technique for secure identity authentication in digital transactions and electronic services. However, reliably distinguishing genuine signatures from skilled forgeries remains a challenging task because of the high intra-writer variability of genuine signatures and the strong similarity between genuine and forged handwriting patterns. This paper proposes an Empirical Mode Decomposition-Driven Deep Neural Network (EMD-DNN) framework for online handwritten signature verification. The novelty of the proposed approach lies in integrating adaptive EMD-based signal decomposition, where the first Intrinsic Mode Function (IMF₁) is exploited as a compact and discriminative representation of handwriting dynamics instead of directly learning from raw signature trajectories. The x-coordinate, y-coordinate, and pressure signals are decomposed using EMD, and the extracted IMF₁ components are combined to construct an informative three-channel feature representation for DNN-based classification. The proposed framework is evaluated on the SVC2004 and MCYT-100 benchmark datasets using a comprehensive evaluation protocol that includes 5-fold cross-validation and quantitative performance assessment based on Accuracy, Precision, Recall, and F1-score. Experimental results demonstrate excellent verification performance, achieving 99% Accuracy, Precision, Recall, and F1-score on the SVC2004 dataset, and 99% Accuracy with 97% Precision, Recall, and F1-score on the MCYT-100 dataset. These results demonstrate that the proposed EMD-DNN framework provides a favorable balance between verification accuracy, computational efficiency, and generalization capability.
Abstract: Bacteria are single-celled organisms that enter the body, grow, and release toxins that harm cells, causing sepsis and other diseases. Because bacteria cause various diseases in humans, prompt diagnosis is required to adapt antibiotic medication and prevent disease spread. This study presents a promising device that can distinguish between different types of bacteria commonly found in the blood. Electronic nose technology is now regarded as a quick tool for detecting pathologies based on volatile organic compounds (VOCs). The use of classical bacteriology takes time to give the practitioner or biologist a diagnosis. The bacterial species is detected from VOCs released by bacteria in a few minutes using a multi-sensor system for the detection of VOCs. The goal of this study was to test and identify ten different types of bacteria in blood by an electronic nose. The proposed models achieved accuracies of 96.77%(LSTM) and 98.91%(Bi-LSTM), demonstrating the superiority of Bi-LSTM for bacterial classification.
Abstract: Accurate blood type classification is crucial for safe transfusions and clinical decision-making, yet existing research is limited by the lack of standardized, publicly available datasets for training and evaluating machine learning models. To address this gap, we introduce ABO-BTI (ABO Blood Typing Image), the first open-source dataset dedicated to blood type classification using high-resolution agglutination images. The dataset comprises 144 cases, with 432 images standardized to a resolution of 1280×590 pixels after processing. This study evaluates the effectiveness of deep learning for blood type identification using the ABO-BTI database. Three models, ResNet50, MobileNetV2, and a proposed deep learning architecture, were trained and tested on the dataset to assess its suitability for machine learning applications. The proposed model achieved an accuracy of 96.51%, significantly outperforming MobileNetV2 (12.64%) and ResNet50 (72.41%). Comparative analysis with traditional machine learning methods further demonstrated that deep learning provides competitive performance while reducing reliance on handcrafted feature extraction. These results highlight ABO-BTI as a valuable benchmark for advancing AI-driven blood type classification. The findings also suggest the potential integration of deep learning-based classification into embedded systems for real-time blood typing in point of care and emergency settings. By providing a standardized dataset and demonstrating the viability of deep learning models, this study lays the foundation for future research in automated blood classification, with implications for both clinical applications and AI-driven medical diagnostics.
Abstract: Using the Electrocardiogram (ECG) signals for biometric identification as a biometric trait has become popular today because of their unique characteristics. The primary objective of this research is to design, implement, and test an efficient and accurate ECG-based biometric identification system that leverages the benefits of Empirical Mode Decomposition (EMD) and Gated Recurrent Unit (GRU) neural networks to improve classification performance and robustness. The proposed method comprises three main stages: preprocessing, feature extraction, and classification. We first start by denoising the signal using a 4th-order Butterworth bandpass filter and then normalizing the amplitude of the filtered signal between 0 and 1. Next, we decompose the normalized signal into multiple intrinsic mode functions (IMFs) using EMD. The first two IMFs are retained as features, and each one of them is segmented into 5-second windows with a 4-second overlap to create a new feature vector. A GRU-based neural network is then trained to identify individuals from their ECG signals. We evaluated the proposed method on three publicly available databases: PTB Diagnostic ECG, MIT-BIH Arrhythmia, and ECG-ID. These are among the most widely used datasets in the field of ECG-based biometric identification because of their diverse characteristics and good representation of various types of heart conditions. We achieved accuracy rates of 99.88%, 99.89%, and 96.34% respectively. which outperformed most existing ECG biometric identification methods. Furthermore, we compared our method with other approaches proposed in the literature. The results show that our proposed system is more accurate and robust than these methods. However, it’s important to note that our method relies on single-lead ECG data, which simplifies signal acquisition but may limit its performance in certain scenarios. Our method can be applied to access control systems, e-healthcare, and biometric identification. Future work could explore incorporating additional modalities to enhance accuracy and robustness.
Abstract: Multimodal biometric systems have received substantial interest for their potential to improve identification accuracy and security in a variety of applications. In this paper, we propose a multimodal biometric system that combines electrocardiogram (ECG) and voice modalities for user identification. The ECG signals are processed using a 6th-order Butterworth filter and segmented around R peaks identified by the Pan-Tompkins++ approach. Instantaneous frequencies extracted using the Hilbert transform are utilized as features and fed into a 1D convolution neural network for classification. Similarly, voice signals are preprocessed by removing silent parts and segmented into 300 ms windows with 150 ms overlap. Mel-frequency cepstral coefficients (MFCCs) are extracted and input into a 2D convolution neural network for classification. Both individual modalities achieve high accuracies of 98.39% and 96.26%, respectively. Subsequently, a score-level fusion model, comprising densely connected layers, integrates the classification scores from both modalities, resulting in a remarkable 100% accuracy. The proposed multimodal biometric system demonstrates the efficacy of combining ECG and voice modalities for robust and accurate user identification. This research contributes to advancing multimodal biometric technologies and holds promise for applications requiring secure and reliable authentication systems.
Abstract: This paper presents an optimised convolutional neural network (CNN) for Electrocardiogram (ECG) biometrics, focusing on enhancing efficiency and performance using a quantised CNN model. The research evaluated the model on 10 subjects from the MIT-BIH database. ECG signals were filtered with a 4th-order Butterworth filter (1-40 Hz), and R-peaks were detected using the Pan-Tompkins++ algorithm. Segments around these peaks were windowed into 125 ms frames, and spectrograms were generated via short-time Fourier transform (STFT). These normalised spectrograms were fed into both standard and 8-bit quantised CNN models for biometric identification. The 8-bit quantised model, deployed on an ESP32, achieved 97.90% accuracy, outperforming the original model. It was 77% faster and 64% smaller in size, demonstrating significant improvements in efficiency and performance. The study suggests further exploration of quantised models in ECG biometrics.
Abstract: Traditional methods of authentication, such as keys, passwords, and PIN codes, are increasingly being surpassed by the robustness and security offered by biometric systems. This study presents a system for human identification using electrocardiogram (ECG) signals, evaluated using two deep learning models: gated recurrent units (GRU) and long short-term memory (LSTM). The proposed methodology begins with signal denoising using a bandpass filter, followed by the decomposition of the signal into multiple intrinsic mode functions (IMFs) via a signal processing technique, empirical mode decomposition (EMD). The first two IMFs, after normalization and segmentation, serve as features for the deep learning models. The performance of these models was assessed on three distinct databases: NSRDB, MIT-BIH, and PTB. MIT-BIH contains a mixture of healthy and ill subjects, while the PTB and NSRDB databases contain only healthy subjects. The GRU model achieved accuracies of 98.57%, 98.26%, and 99.17% on these databases respectively, while the LSTM model achieved accuracies of 98.33%, 97.89%, and 98.27% respectively, demonstrating its reliability for biometric applications. While the proposed system was tested only on ECG, its applications can be extended to include other biomedical signals, indicating its potential for broader biometric use.
Abstract: Reliable personal identification is crucial. This study explores an ECG-based system using spectrograms and involutional neural networks (INNs) for accurate person identification. The system first preprocesses ECG signals with a Butterworth filter and segments them using the Pan-Tompkins++ algorithm. Each segment, representing a heartbeat, is then converted into a spectrogram using the short-time Fourier transform (STFT). Finally, an INN classifies the spectrograms for identification. Tested on the MIT-BIH Arrhythmia and Physikalisch-Technische Bundesanstalt (PTB) databases, the system achieved accuracies of 97.93% and 97.63%, respectively, surpassing conventional methods like convolutional neural networks (CNNs). This improvement is attributed to INNs’ ability to better capture both local and temporal patterns in the spectrogram data.
Abstract: Multimodal biometrics employs multiple modalities within a single system to address the limitations of unimodal systems, such as incomplete data acquisition or deliberate fraud, while enhancing recognition accuracy. This study explores score normalization and its impact on system performance. To fuse scores effectively, prior normalization is necessary, followed by a weighted sum fusion technique that aligns impostor and genuine scores within a common range. Experiments conducted on three biometric databases demonstrate the promising efficacy of the proposed approach, particularly when combined with Empirical Modal Decomposition (EMD). The fusion system exhibits strong performance, with the best outcome achieved by merging the online signature and fingerprint modalities, resulting in a normalized Min-Max score-based Equal Error Rate (EER) of 1.69%.
Abstract: Multimodal biometrics is the technique of using multiple modalities on a single system. This allows us to overcome the limitations of unimodal systems, such as the inability to acquire data from certain individuals or intentional fraud, while improving recognition performance. In this paper, a study of score normalization and its impact on the performance of the system is performed. The fusion of scores requires prior normalisation before applying a weighted sum fusion that separates impostor and genuine scores into a common interval with close ranges. The experiments were carried out on three biometric databases. The results show that the proposed strategy performs very encouragingly, especially in combination with Empirical Modal Decomposition (EMD). The proposed fusion system shows good performance. The best result is obtained by merging the globality online signature and fingerprint where an EER of 1.69% is obtained by normalizing the scores according to the Min-Max method.
Abstract: The demand for reliable identification systems has grown recently. Using the mean frequency, median frequency, band power, and Welch power spectral density (PSD) of ECG data, we proposed a novel biometric approach in this study. ECG signals are more secure than other traditional biometric modalities because they are impossible to forge and duplicate. Three different support vector machine classifiers—linear SVM, quadratic SVM, and cubic SVM—are employed for the classification. The MIT-BIH arrhythmia database is used to evaluate the suggested method’s precision. For the linear SVM, quadratic SVM, and cubic SVM, respectively, test accuracy of 93.6%, 96.4%, and 97.0% was obtained.
Abstract: Biometric authentication systems have become indispensable in modern security infrastructures, supporting reliable identity verification across domains such as access control, forensic analysis, healthcare, and secure digital services. In this context, the present survey examines hidden biometric modalities, an emerging and highly promising frontier in authentication research. Unlike conventional biometric traits such as fingerprints, facial appearance, or iris patterns, which are often visible, externally captured, or relatively easy to imitate, hidden biometrics are derived from physiological or behavioral signals that are less accessible to adversaries and therefore offer stronger protection against spoofing and impersonation. This survey provides an integrative review of major hidden biometric modalities, including deoxyribonucleic acid (DNA), electrocardiography (ECG), electroencephalography (EEG), and electromyography (EMG), as well as medical imaging-based approaches such as magnetic resonance imaging (MRI) and X-ray imaging. A central contribution of this work is the comprehensive examination of the full biometric pipeline across these modalities, from data acquisition protocols and sensor configurations to preprocessing strategies, feature extraction methods, and classification algorithms. In particular, we analyze the strengths and limitations of traditional signal-processing techniques and modern machine-learning approaches, highlighting how methodological choices influence recognition accuracy, robustness, and practical deployment. To enable a structured comparison, the survey evaluates each modality according to widely accepted biometric criteria, including universality, uniqueness, permanence, circumvention resistance, and overall performance. This comparative perspective makes it possible to identify both the operational advantages and the technical challenges associated with each hidden biometric source. The novelty of this survey lies in its unified treatment of signal-based and image-based hidden biometric modalities within a single conceptual framework. By bridging these two research streams, the paper offers a comprehensive reference for experienced researchers while also serving as an accessible entry point for scholars new to this interdisciplinary field. Ultimately, the survey aims to clarify current progress, expose unresolved challenges, and outline future research directions toward secure, reliable, and privacy-preserving authentication systems.
Abstract: This research paper introduces a novel Convolutional Neural Network (CNN) architecture optimized for speaker recognition on resource-constrained devices. The proposed model was evaluated on a subset of 10 speakers from the LibriSpeech database. To process the audio signals, preprocessing techniques to remove silent parts and segmented the signals into 1-second windows with a 500 ms overlap were employed. The Mel Frequency Cepstral Coefficients (MFCC) of each window were then extracted and fed into the CNN model. After training, the CNN model was quantized, converted to TensorFlow Lite (TFLite) format, and integrated into an Arduino-compatible library. This deployment was successfully executed on the ESP32 board. The proposed quantized model achieved an impressive accuracy of 98.58%, outperforming the original model by being 53% faster and 73% smaller in size. The findings of this study prove the efficiency of deploying accurate and fast speaker recognition systems on edge devices.
Abstract: This paper presents a novel biometric system based on electrocardiogram (ECG) signals and tiny machine learning (TinyML) techniques for personal identification. A subset of 10 individuals from the MIT-BIH Arrhythmia Database was used to develop and evaluate the proposed system. In the preprocessing stage, the ECG signals were denoised using a 4th-order Butterworth bandpass filter. Subsequently, the signals were segmented into 100ms windows centred around each r-peak, which were detected using the Pan-Tompkins++ algorithm. To ensure uniformity, the amplitudes of the segmented windows were normalized within the range of 0 to 1. For the classification, a deep learning model consisting of an input layer and three fully connected layers was employed. To enable deployment on resource-constrained devices, the trained model was converted to TensorFlow Lite format and further transformed into an Arduino library.
Abstract: Biometric systems that provide accurate identification are essential for making informed decisions in various fields. In this study, we propose a novel approach for identifying medical patients using the QRS waveform of their ECG signal. ECG signals are a highly secure biometric trait as they are difficult to spoof and reproduce. To derive a representative QRS waveform, the signals were subject to preprocessing via a 4th-order bandpass filter, amplitude normalization, and R-peak detection through the Pan-Tompkins algorithm. Subsequently, a vector encompassing 50 samples before and after each index was generated. An identification accuracy of 99.47% was achieved using 48 healthy subjects from the PTB Diagnostic ECG Database. To further test the system, we used 282 healthy and diseased subjects from the same database and achieved an accuracy of 97.60%. Our proposed approach provides a promising solution for biometric identification, which can improve decision-making in various fields such as security and healthcare. With the increasing demand for accurate identification, our study highlights the potential for ECG-based biometrics as a secure and reliable modality for decision-making.
Abstract: This paper presents an edge-based online signature identification system using Tiny Machine Learning (TinyML) and an ESP32 microcontroller. The proposed system utilizes the SVC2004-task1 dataset and applies Gaussian filtering, position normalization, size normalization, and length normalization as preprocessing techniques. A deep learning model with three fully connected layers is trained using Edge Impulse framework, and then deployed on the ESP32 board after quantizing the model into 8-bit. The system achieves a high identification accuracy of 94.94%, demonstrating the feasibility and effectiveness of using TinyML and edge computing for online signature identification, which has potential applications in security and identification systems.
Abstract: In this paper, we propose an ECG biometric system based on instantaneous frequencies extracted by the Hilbert-Huang Transform (HHT). The system was evaluated using two different databases: PTB Diagnostic ECG and MIT-BIH Arrhythmia. The results demonstrate the robustness of our system, as we achieved a testing accuracy of 95.31% for the MIT-BIH database and 96.42% for the PTB database. For the classification, we used a deep learning model based on 3 gated recurrent units (GRU) cells. This paper aims to verify the efficacy of using instantaneous frequencies as a feature for biometric identification.
Abstract: The emerging long short-term memory neural networks (LSTM) are the widely used type of recurrent neural networks (RNN). A novel ECG biometric method based on the ensemble empirical mode decomposition (EEMD) and instantaneous frequencies extracted by the Hilbert-Huang transform (HHT) is presented. Deep learning and bidirectional long shortterm memory neural networks were used for classification. The performance of the system was evaluated using PTB Diagnostic ECG Database. A testing accuracy of 89.33% was achieved when the evaluation was done with the full database and an accuracy of 97.42% is achieved when the evaluation was done with 5 subjects. The main goal of this paper is to confirm the effectiveness of using instantaneous frequencies as a feature for biometric identification.