With increasing popularity of wearable and Body Sensor Network (BSN) technologies, there is a
growing concern on the security and data protection of such low-power ubiquitous devices. With
very limited computational power, BSN sensors often cannot provide the necessary data protection
on the sensitive personal health information they collect and process. Biometrics, such as face and
fingerprint, have been widely used for securing computer systems and mobile devices, however, such
methods have issues. For instance, the capturing of the biometric is quite intrusive and previously
collected data or compromised data can be reused by attackers.
The aim of the thesis is to tackle the challenges of collecting biometrics pervasively with miniaturised
BSN nodes, and ensuring the data freshness of a BSN security system, by investigating innovative
ways of using behavioural biometrics. It is hypothesised that behavioural biometrics, such as Electroencephalographic (EEG) and walking patterns (gait) can be used for unobtrusive encryption of
BSN wireless communication channels and secure the BSN-based healthcare systems.
A person’s brain wave signal, also known as EEG signal, is nearly impossible to mimic and can be
easily collected with EEG headsets without user intervention; therefore, it is suitable to be used as
biometrics for securing BSNs. Due to the complex nature of EEG signals, the state-of-the-art manually feature extraction methods often cannot utilise the full potential of the underlying features neural
activities in the EEG signals. Therefore, to explore the potential of using EEG for securing BSNbased healthcare systems and to improve the performance of the current EEG-based authentication
systems, the use of deep learning approaches is investigated.
Although EEG-based security systems perform exceptionally well, EEG headsets are still very expensive and cumbersome in size. To reduce the costs of the security systems, the walking pattern
of a person, called gait, is investigated as a biometric for securing BSNs. Gait is one of the most
promising behavioural biometric traits for securing wireless communications between BSN sensors
and coordinators. This thesis presents the work in resolving issues in using gait for BSNs, especially
the challenge of using less correlated gait signals collected from sensors located at different positions for the common entropy sources of Biometric Cryptosystems (BCS). In this context, a novel
light-weight symmetric key generation scheme based on the timing information of gait and fuzzy
commitment scheme is proposed. The effect of gait-based soft biometrics is also investigated, namely
age and gender. Through analysing a large gait database with inertial sensor data, and the results show that age and gender information can be accurately estimated using only gait signals. The recognised
age and gender information can be used to improve current gait-based security systems.
Next, with the advances of Artificial Intelligence (AI) for the healthcare applications, many wearable
devices and BSN sensors are now able to perform on-node inferencing. The challenge of less correlation between gait signals at different body positions can be tackled by machine learning techniques.
An Artificial Neural Network (ANN) framework has been developed for estimating gait signals on
the shin and thigh positions from gait signals collected on the ankles. The work shows the possibility of using ANNs to project gait signals captured from one body position to onto another position.
Therefore, based on this finding, the ANN framework is proposed to improve the previously proposed
gait-based key generation scheme, where gait signals collected on the head, upperarm, wrist, waist,
thigh, and shin positions are all projected onto the chest position so that the transformed signals are
highly correlated and more similar secret keys can be extracted by devices at those positions.
From our experiments on using biometrics for securing BSNs, it is also found that the freshness
of gait signals can be used to generate random numbers by removing the low frequency periodical
components in the signals. The last part of the thesis investigate the use of gait signals as the entropy
source for Random Number Generators (RNG), and a novel random number generation method is
proposed for securing on-body IoT devices based on temporal signal variations of the outputs of
the Inertial Measurement Units (IMU) worn by the users while walking. The proposed method has
been tested with two inertial gait datasets and passed four well-known randomness test suites, namely
NIST-STS, ENT, Dieharder, and RaBiGeTe.