ECG-BASED HUMAN IDENTIFICATION AND AUTHENTICATION

Abstract

Identity theft is a major concern today. It has serious ramifications beyond data and personal information loss, necessitating the implementation of robust and efficient user identification systems. Identification and authentication of users is foremost for granting access to personal devices, accounts or confidential files. Traditional techniques include PINs, passwords, and tokens, which are viable to be forgotten, stolen or cracked. A fundamental development in the field of identity theft protection is biometrics, which is a reliable alternative to classical identification and authentication approaches. Of all the biometrics, the Electrocardiogram (ECG) has been on a roll due to its inevitably desirable intrinsic features that are not available in other biometric traits. Unlike other biometrics, it cannot be copied (signature), imitated (voice), duplicated (finger prints, iris etc.) and faked (gait). Moreover, it is the proof of aliveness of the person. ECG biometrics identification has been implemented using time domain, frequency domain, hybrid of the two, and very recently, the Deep Learning (DL) approaches. Recent state-of-the-art techniques involve CNNs for ECG based user identification. However, the CNN based methodologies have limitations of long training time and large datasets. The current research work is divided into three stages. First, a dataset of diverse ECG waveforms was created. The dataset was carefully built to include recordings from both genders, a wide age range, various signal-acquiring setups, different health and physical conditions, various physical postures and mental conditions. The signals in the dataset were prepared to be used with deep learning algorithms. This involved normalizing the amplitude and period, filtering, segmenting, and ultimately converting them into images. Second, several deep learning algorithms were explored and experimented using the prepared dataset, and the results were analysed. Two novel DL models have also been developed that address the objectives of increased identification rate (IDR) and decreased computational effort. Third, a cancellable ECG-biometric has been designed for secured ECG-based human authentication. The main aim of this study is to create an automated biometric system for user identification and authentication using ECG waveforms. In this investigation, an initial examination encompassed four distinct deep learning (DL) algorithms for the purpose of user identification utilizing the provided dataset. These DL algorithms comprise an LSTM network, a transfer learning approach employing ResNet-50, a customized residual network, and a customized dense network. Subsequent analysis of the findings revealed the dense network to be the most suitable for the given task. Extensive investigations were conducted on the Dense network for four distinct architectures, while meticulously analysing the activations of each layer. Furthermore, experiments were undertaken to assess the impact of reduced training dataset and multiple session ECG recordings on the network's performance. The current study also developed two deep learning (DL) models for user identification based on ECG data. The first model is an ensemble of UNet with ANN. The UNet part of the model is specifically designed for segmenting and extracting crucial features from the ECG waveform, while the ANN is iii

iv

responsible for classification. This model has also been implemented using TensorFlow/Keras and is suitable for deployment on portable devices with limited capacity. The proposed ensemble model generates a considerable number of learnable parameters, leading to escalated computation costs and enhanced hardware requirements. Consequently, a hybrid model comprising a dense CNN with a fire module was formulated. The implementation of the fire module within the network resulted in reduction of the learnable parameters by a considerable factor, thereby decreasing the computational effort, memory demand, and training duration. However, this reduction in learnable parameters has an impact on the model's performance. A study has also been conducted to develop a human authentication system that takes into account hardware limitations and computation costs. A simple yet effective ECG secret key generation technique for user authentication has been developed based on multiple recursive algorithm (MRA). The proposed keys are generated at two stages, with one part stored in the device and the other part generated on-the-fly. Concatenating both parts generates a 128-bit secret key that can be used for authentication. The keys generated with this technique are evaluated using metrics such as reliability, robustness, and entropy. The analysis of the results reveals that the proposed key generation technique can provide sufficient strength and randomness to the keys, ensuring a secure ECG-based user authentication system.