Low-cost fitness and activity trackers for biometric authentication

Abstract

Fitness and activity tracking devices acquire, process and store rich behavioural data that are consumed by the end-user to learn health insights. This rich data source also enables a secondary use of being part of a biometric authentication system. However, there are many open research challenges with the use of data generated by fitness and activity trackers as a biometric source. In this article, the challenge of using data acquired from low-cost devices is tackled. This includes investigating how to best partition the data to deduce repeatable behavioural traits, while maximizing the uniqueness between participant datasets. In this exploratory research, 3 months’ worth of data (heart rate, step count and sleep) for five participants is acquired and utilized in its raw form from low-cost devices. It is established that dividing the data into 14-h segments is deemed the most suitable based on measuring coefficients of variance. Several supervised machine learning algorithms are then applied where the performance is evaluated by six metrics to demonstrate the potential of employing this data source in biometric-based security systems.

Introduction

The use of fitness and activity tracking devices is rapidly increasing.¹ They provide the capability to track activity related to fitness and health, providing the user with essential health insights as well as enabling them to reach their fitness goals. There is a large array of different devices available on the market, including high-end devices manufactured by Apple, Fitbit, Garmin and Samsung. Low-cost alternatives are available that focus solely on fitness tracking from manufacturers such as Xiaomi Mi, Honor and Amazfit. Each device has different capabilities [1, 2], and manufacturers are striving to increase the range and quality of sensed data [3], as well as unique ways to process, visualize and provide insights to differentiate their products. This is resulting in the devices becoming popular and ever appealing to the consumers who are seeking reliable and high-quality fitness tracking functionality. In this work, we investigate the secondary use of this rich data source, which is the potential of using it in a biometric system for authentication purposes.

Biometric research is an active discipline with researchers identifying and developing new biometric systems. For example, in recent work, vein patterns are used with a high-level of success [4]. Other recently developed biometric systems range from smartphone behavioural patterns [5] to knuckle biometrics [6]. Biometrics based on physical characteristics of the human body demonstrate a good degree of suitability due to their uniqueness. Although face, fingerprint, voice, eyes and other physical biometric systems are already in operation, a fundamental limitation of such systems is that the data is more challenging to acquire and is often invasive to the user [7, 8]. For example, many high-resolution and close-up images are required for training a facial recognition system. Furthermore, several external conditions need to be satisfied regarding the angle, distance, lighting, etc. for the authentication process [9]. Therefore, researchers have been probing the feasibility of behavioural biometrics, which is an active research area, mainly due to the extensive usage of wearable fitness devices and availability of the required data. Behavioural biometrics are much more appealing as the users do not need to use invasive technologies to acquire data samples for authenticating participants.

In this research, we focus on exploring fitness activity data as a behavioural biometric without the need for invasive sampling. This is by no means a new concept and other researchers have previously discussed the potential [10]; however, a key aspect to this study is that low-cost hardware devices are used to investigate the real-world applicability of this approach. Furthermore, numeric biometric data of 3 months from the smartwatch have been utilized for training and testing the machine learning algorithms. Another key aspect of this study is the use of raw biometric data in a time-window based approach, gaining promising levels of classification accuracy (ACC) without applying any complex and computationally intensive feature extraction technique. Many previous works employ intensive feature extraction techniques to improve data representation, and consequently the accuracy, such as shown in existing works by Blasco and Peris-Lopez [11] and Hira and Gillies [12]. Therefore, our solution provides a mechanism to improve efficiency, that is, reduced computation cost due to evading the feature extraction process, yielding at worst a 15% decrease in accuracy. Furthermore, in previous research, the authors do not consider the implication of using different machine learning algorithms, and more importantly, their datasets are not publicly available for use.

The fundamental approach is that a user would acquire fitness health data over a pre-defined duration, which can subsequently be used to establish a routine pattern and enrolled within the biometric system for future matching. In this work, enrolment is performed in terms of using supervised machine learning to learn how to classify different users. However, many challenges need to be eliminated and will be addressed in this work. The first is to determine how historical data can best be divided, allowing machine learning algorithms to distinguish the participants accurately. The second challenge is to establish which supervised machine learning algorithms are most suitable for processing data of this type. The overall research question has been addressed in this work is based on empirical observations, can fitness health data acquired from low-cost hardware be used as a biometric?

In this article, we present the following contributions:

• the use of the raw minute data values for steps, heart rate, and sleep to establish an optimal time-windowing technique to determine how much historical data is required in the subset used to implement an intrinsic biometric system, rather than identifying features across the multiple data sources and

• the use of a dataset consisting of 3 months’ worth of data for five anonymized participants, which we have made available.² Previous studies may have used data for a higher number of participants and longer duration; however, their data is not yet publicly available.

The article is structured as follows: first, a discussion of related work followed by information on the specific device used and participants of this study is provided. Following on, we present an example of the acquired data, explaining how we convert the data to a standard frequency of 1 min. Next, we explain our time-windowing technique and perform empirical analysis to determine the optimal time-window in regards to maximizing the uniqueness among participants’ biometric data. This includes improving the identification of repetitive patterns, which could be due to the similar physical activities in the participants’ day-to-day routine. Finally, we investigate and analyse the use of many different supervised machine learning algorithms to establish the potential of training and using machine learning to match participant data. A conclusion is then provided, presenting future work in this direction.

Related Work

Components of a smartwatch

Wearable products provide useful services to their owners and help in leading a better lifestyle. Specifically, the wrist worn devices enable original equipment manufacturers to deliver wellness and fitness related services, which is a crucial reason behind these devices gaining popularity. Key sensors such as accelerometer, gyroscope, magnetometer and pressure sensors are essential for smartwatches and activity trackers to capture the data necessary for these services. In terms of the specific technology for acquiring health-related data, the Mi band 3 is utilized and consists of four major sensors: triple axis accelerometer (TAA), photoplethysmography (PPG), proximity (p-sensor) and actigraphy, according to the official specification.³

The TAA sensor is used to detect and count the steps, both during walking and running, along with measuring the distance travelled, speed and calories burned [13]. It is more accurate than a pedometer as it describes any motion with the help of three components: forward (x-axis), vertical (y-axis) and side (z-axis) as shown in Fig. 1b. The sensor can recognize both static (gravity/g-sensor) and dynamic (vibrations and movement) accelerations. The TAA sensors emit corresponding electrical charge under stress, intense activity or force of movement, which is converted into an acceleration value.

Figure 1:

(A) A photoplethysmography (PPG) sensor. The PPG sensor determines the heartbeats per minutes using reflected light. ‘T’ is the transmitter of light, where ‘R’ is receiver of reflected light. (B) A triple axis accelerometer. Shows x, y and z axis of measurement in a triple axis accelerometer

The PPG sensor is used to determine heartbeats per minute, which is based on detecting blood volume changes in the microvascular bed of tissues [14]. It is a simple, low-cost, and portable alternative of electrocardiogram (ECG) that uses optical technology. Because of the small form factor, the PPG sensors are easily embedded in wearable devices. The sensors consist of a transmitter and a receiver along with an optical barrier in the middle to prevent sending direct light to the receiver, as shown in Fig. 1a. The transmitter emits visible light onto the surface of the skin from a moderate distance using light emitting diodes. Human blood is good at absorbing green light and also reflecting red light. With every heartbeat, the capillaries expand and contract because of the change in blood volume. The receiver absorbs the reflected light, measuring blood flow as per the intensity of light. Using this data, it is possible to measure heartbeats per minute [15].

Actigraphy and p-sensor are used to measure the sleep cycle [16]. The actigraphy tracks the overall sleep time, which includes deep sleep, light sleep and awake time during sleep. The state of sleep is commonly identified by the lack of movement of the wrist worn device using an accelerometer [17]. For this purpose, a sensor is used called actimetric, which records and measures the body movements (both subtle and intense), and the output data are plotted on a graph and analysed to determine the sleep cycle. The additional p-sensor is used to identify the presence of skin. This means if the smartwatch is sitting idle on a table, that period would not be considered as a sleep, hence providing an accurate sleep measurement.

Using machine learning in biometric systems

There are many instances where biometric systems have been built upon supervised machine learning approaches. A recent study used multi-layer perceptron (MLP) and radial basis function (RBF) neural networks to perform classification of ECG signals [18]. The MLP is a supervised linear classification technique that uses a feed-forward neural network made up of input, hidden and output layers. Similarly, the RBF determines the inputs’ similarity within the training set to perform classification. Both techniques were tested on 324 samples and presented 98% accuracy for MLP and 97% for RBF. A similar paper presented the use of restricted Boltzman machines and deep belief networks to classify and recognize PPG signals [19]. The signals are first filtered to reduce noise and then segmented to determine the distinguishing features. Interpolation and extrapolation are also used to ensure that all segments are of the same length. Following on, features are extracted that includes mean, standard deviation, minimum/maximum value of a segment, etc. The experimental results are performed on 12 subjects and have shown 96.1% accuracy using 10-fold cross-validation technique.

In other research, the authors proposed the use of a neural network to perform the classification of normalized ECG signals for user authentication [20]. This was tested on 90 individuals and presented 99.54% accuracy. Another research work proposed an authentication scheme using ECG biometrics, which extracts and stores signal features of each individual in the database, and later use a statistical matching threshold for authentication [21]. This solution demonstrated 90% accuracy on 73 individuals. A research study claims that the heart sound signal can also be used for biometric authentication. In previous research, the authors combine heartbeat and acoustic signals to recognize the identity of individuals [22]. The heartbeat signals were collected from a Vernier sensor, and Littmann Electronic Stethoscope Model 4100 WS was used for heartbeats. To extract features, the approach uses the entire ECG signal for heartbeats and the short-time Fourier transform along with the Discrete Cosine Transform (DCT) coefficients to process heart acoustics. The developed solution was tested on 21 individuals and achieved an accuracy of 97%. Other researchers employed the use of a support vector machine (SVM) classifier for the identification and authentication of subjects using minute-level biometrics [23]. The dataset used in this article was collected from the undergraduate students of the University of Notre Dame. The original dataset, named as NetHealth [24], contained approximately 700 subjects but the paper considered and analysed 421 subjects. This is one of the large-scale studies that also extracted 108 (sedentary) and 109 (non-sedentary) features from sleep status, step count, heartbeat, calorie burn and the metabolic equivalent of task. The results have shown 0.93 and 0.90% accuracy for sedentary and non-sedentary periods, respectively.

A research study utilized actigraphy, which is a non-invasive monitoring technique, for the identification of sleep–wake states within infants [25]. The authors employed statistical and neural network-based algorithms and showed an accuracy between 77 and 92%. It is claimed that actigraphy is a comparable alternative to the similar polysomnography technique, which is difficult to use and intrusive in nature. A recent patent proposed the use of sleep physiology for authentication of subjects by creating biometric profiles in order to either verify or disprove the identity of the individual [26]. The patent suggests the collection during the different stages of sleep include heart rate, heart rate variability, acceleration and breathing rate. Using this data, a biometric profile can be created for each subject. Following on, classification algorithms, such as random forest, logistic regression or neural network can be used for subject identification. Another recent patent suggests the use of a device that can be used for both medical diagnosis and identity verification [27]. This is performed with the help of multiple sensors attached to the patient (or bed in the case of sleep analysis) that are constantly collecting data and authenticating the patients.

Several publications propose the use of footstep and gait-based biometrics for authentication [28]. The footstep signals are generated at 1024 Hz frequency through a rubber floor containing pressure sensors. The signals are based on varying voltage output due to the different amounts of pressure placed on the sensors. Various machine learning techniques, such as SVMs and deep neural networks are then used to perform identification and authentication. Another paper proposed a gait recognition technique for user authentication on mobile phones [29]. An important contribution of this research is the use of built-in accelerometer of mobile devices, which is placed around the hip location for signal collection purposes. It uses dynamic time warping as a distance metric between stored and current signals and achieved an accuracy of 80% over 51 individuals. Another article [30] explored the use of Hidden Markov Models in gait recognition using similar smartphone accelerometer data and a seven state Hidden Markov Model with five Gaussian mixtures. This method achieved an accuracy of 90% using data from 48 participants. A recent study proposed the use of convolution neural networks to recognize footsteps in a floor sensor system for biometric application [31]. Using a dataset containing 20 000 samples from 127 subjects, the solution achieved an equal error rate (EER) of 9–13%, which is claimed to be a 4% performance improvement over the existing solutions. Another recent paper presents an interesting system called AcousticID [32]. This system utilizes a SVM to classify acoustic signals generated by off the shelf devices to identify users. The performance of AcousticID was assessed using 50 participants in an area of 60 m², and achieved an average accuracy of 96.6%.

Previous work has demonstrated the potential of using supervised learning within biometric systems processing health-based data sources. Although previous studies have shown good accuracy values with large participant numbers, there are limitations over the data collection with regards to cost, use of specialist equipment and requirement of large experimental setups [33]. In this article, we have used a low-cost, consumer-ready product for the participant identification. Furthermore, one of the main aims of this study is to determine the strength and feasibility of acquired biometric data, unlike existing schemes that involve computationally intensive feature extraction techniques. This will help in exploring the real-world utilization of proposed solution. In addition, we examine and discuss the suitability of several machine learning algorithms. All steps, including data elicitation, analysis and utilization, have low development overheads and have shown promising accuracy.

Device and participants

In this work, we utilize biometric data generated by the MI band 3 device. The MI band 3 records data in three key areas: (i) heart rate; (ii) step count and (iii) sleep. We use the MI band 3 due to the ease of data extraction, its low cost, wide availability, useful features and also the long battery life. All of these benefits also demonstrate the feasibility of using relatively cheaper hardware in real-time biometric solutions. The device is built on Dialog’s DA14681 System-on-Chip, which offers developers greater processing power, resources, range and battery life for a wide variety of connected consumer applications, including high-end fitness trackers and other wearable devices such as the new Samsung Galaxy Fit. It may well be that the sensing capabilities are less accurate and reliable than other alternatives; however, as this is exploratory study, it makes sense to use data that are generated from low-cost devices as it is more indicative of the data generated by real users and will inevitably provide greater variation. It is also worth noting that the MI Band 3 analysis software is capable of providing analytic capabilities and generating secondary data sources, such as calories burned; however, in this study we stick to primary data sources of heart rate, steps and sleep.

We use data generated from five participants, each wearing a different MI band 3 for a 3-month period. Each participant wore the watch 24 h a day, apart from when the battery needed to be charged (Mi band 3 can be fully charged in 75 min), which was around once per 2 weeks. The participants are all male and work in an office environment. They are all aged between 20 and 30 years of age. Further biographic information is not available for the participants as we are making the data publicly available for the research community. The participants do not wish to disclose their identity, although are willing to share their data anonymously. It is worth mentioning that as they are all employed in similar job roles and of a similar age, there is potential for their data to be quite similar, resulting in lower accuracy. However, as we are performing an exploratory study, using data that might lack significant difference is more realistic of the real-world and selectively choosing data from individuals’ with significantly different lifestyles could introduce bias.

Data analysis

As previously mentioned, when considering the use of health data sources for biometric purposes, it is important to consider how the data are to be pre-processed to account for missing values, as well as how to partition the data for machine learning application.

Data pre-processing

The first stage is to pre-process the data, converting it from its original form into a structure suitable for processing. The data sources processed in this research are that of heart rate, step counts and sleep data. However, to ensure they are suitable for this research, the process described in this section has been devised. It is important to note that a time-unit of 1 min is selected for the acquired data as the device only records and exports data at the minute frequency.

Heart rate

Heart rate is currently recorded by the device at predefined intervals (typically every 5 min) or when it detects activity and increases frequency to acquire fine-grained information when the user is performing exercise. This results in a non-uniform time distribution due to the unpredictability of exogenous user activity. As previously mentioned, we adopt a fixed frequency of recording the data once per minute; however, in some instances, there is an absence of data recorded. To overcome this issue, linear interpolation (LI) and linear extrapolation (LE) techniques [34] are used to determine missing recordings on a minute-by-minute basis. The LI constructs missing data points within the dataset. For example, if a subject does not wear the watch for 75 min due to charging a low battery, we applied LI to find the missing recordings for this particular time period. On the other hand, the LE estimates heart rate values beyond the original dataset. For example, if a subject took off the smartwatch an hour before the designated time on the final day of data collection. We then utilized LE to determine such missing recordings, hence completing data for a full 24 h of the day. Both LI and LE are required in this solution as they allow us to provide the same amount of recordings against each subject (i.e. balanced datasets), preventing classifier bias issues [35]. Figure 2a illustrates an example whereby the discrete data values recorded by the watch can be seen as the blue circles, and the interpolated and extrapolated values are represented by the red ‘+’ sign.

Figure 2:

Heart rate, steps and sleep plots over a 24-h period. (A) Heart rate over a 24-h period, including interpolated and extrapolated data points. (B) Steps over 24-h period. Note that zero values are omitted from the graph to display for the ease of viewing recorded data. (C) Sleep over a 24-h period, displaying minutes awake, light sleep and deep sleep as a binary value with zero values are omitted for ease of viewing

Step data

Step data are recorded by the device in 1 min segments, but only when the participant is moving. Therefore, it is assumed that when there is no value recorded, the participant is not moving. Figure 2b illustrates a participant’s 24-h step activity and it is evident that there are times in the day where the user is not moving as movement is taking place between approximately 09:45 and 22:00. For the ease of viewing the graph, zero values have been removed; however, it is worth noting that in the dataset, every value will have a recording of zero, if no steps were recorded.

Sleep

Sleep is recorded as a sequenced string of light (L), deep (D) and awake (A) sleep minutes (e.g. 10L28D14L36D2L9D). In this research, the string is separated into three binary values to represent whether the individual is awake, light or deep sleeping at the specific minute. Figure 2(c) illustrates sleep data for one night, between 23:58 and 07:10. The same applies as with step data where zero values are omitted from the graph as they represent the duration when the user is awake. It is also worth noting that a true value recording against ‘awake’ means that the participant is awake after initially falling asleep and before falling asleep once again. It does not represent all the minutes that the participant is awake during the day, only within their sleep cycle.

Time-window selection

After performing initial analysis on pre-processed data with a 1-min sampling rate, it became apparent that not only are there too many data points resulting in a decrease in algorithm performance, but there is also an insufficient difference in the individual 1 min’s worth of data to differentiate between the participants adequately.

The next stage, therefore, is to explore different time-window duration to determine which has a higher degree of uniqueness between participants. The justification for exploring this is to understand how we should partition the data for use to train machine learning algorithms. Furthermore, as we are considering the use of the data sources for biometric purposes, it is important to gain an understanding of how much data (in time length) will be required for accurate matching. More specifically, what is the optimal time-window of data that maximize the ability for a correct match to be made. It is worth noting here that for the purposes of authentication, the historic time-window segment would be extracted from the device. For example, if a time-window of 1 h was used, then the values acquired over the previous hour will be extracted and used to authenticate an individual. This is necessary to avoid the user being inconvenienced during authentication and also to capture the user’s normal behaviour. Collecting data for the purpose of authentication might result in them being inactive while the data are acquired.

It is necessary to systematically process the acquired data to determine the most appropriate time-window duration. Here, a time-window refers to a subset of the data to a specific length, for example 30 min. Our motivation for starting with a minimum of a 30-min time-window is that the data are recorded by the device on the minute level, and the three values (heart rate, steps and sleep) for a 1-min period are going to be insufficient to adequately represent uniqueness among participants. This is because the data represent the participant’s behavioural activity and, if the time-window is too small, not enough of the activity will be captured in the time-window dataset. The remainder of this section presents the undertaken process and analysis results. To partition the data, the following process is performed:

The data are divided into subsets based on the time-window’s duration. In this research, we consider a duration of 30–1440 min (24 h) in 30 min increments. Once the five datasets have been created, we then calculate the five following values: mean heart rate, aggregated step count, total awake minutes, total light sleep minutes and total deep sleep minutes for each time-window segments. More specifically, the time-window data will be collapsed into one single mean value for each source. Following on, we calculate the coefficient of variance (CV) for each time-window segment using the following equation. The approach of utilizing a CV measure is a widely adopted process when quantifying the variation within a dataset, as well as the difference between datasets. In this work, it is particularly important to measure the variance to measure how well the time-window data contains similar values, which could be representative of the participants repeated and distinct behaviour. The process adopted in this research is inspired by works undertaken in medical research disciplines with similar data processing requirements [36–38].

$$\text{CV}=\frac{\sigma }{\mu }$$

(1)

Where σ is the standard deviation and μ is the mean. We calculate CV values for each of the five datasets (heart rate, steps, etc.). Next, the standard deviation is calculated using the following equation:

$$\sigma =\sqrt{\sum _{i=1}^n{(x_i-\mu )}^2}.$$

(2)

Where n is the number of data samples in the dataset and x_i is an individual data value. This process is repeated for each of the five datasets. Once we have the five CV values, $C=\left\{{\text{CV}}_{\text{heart}}{,\text{CV}}_{\text{steps}}{,\text{CV}}_{\text{deep}}{,\text{CV}}_{\text{light}}{,\text{CV}}_{\text{awake}}\right\}$⁠, we create a combined measure for the time-window (e.g. CV₃₀ for 30 min) by using the following two equations, whereby the values are first normalized and then summed. The first equation normalizes the data using min–max normalization [39]:

$${\text{CV}}_n=\frac{{\text{CV}}_n-\text{min}(C)}{\text{max}(C)-\text{min}(C)}$$

(3)

Where CV_n is one of the five CV values (e.g. CV_heart). The set CN is used to store the five normalized CV values The ${\text{CV}}_{\Delta T}$ value is then used to store the summed CV_n values, where $\Delta T$ is the time-window duration. The following equation creates the combined variance measure for the 30 min time-window (i.e. $\Delta T$ = 30 min):

$${\text{CV}}_{30}=\sum _{i=1}^{|C|}{C}_i$$

(4)

This process is then repeated for each participant, generating a ${\text{CV}}_{\Delta T,p}$ for each, where p is the participant number and $\Delta T$ is the time-window duration. The set CV_p is introduced to store these values, $\text{CV}p=\{{\text{CV}}_{\Delta T,1},{\text{CV}}_{\Delta T,2},\dots ,{\text{CV}}_{\Delta T,p}\}$⁠. For example, ${\text{CV}}_{\Delta 30,1}$ would represent the combined variance measure for the 30-min time-window for participant 1.

A measure of variance in CV_p is acquired by calculating $d_{\Delta T}=\text{max}(\text{CV}p)-\text{min}(\text{CV}p)$⁠. The set D is used to store all $d_{\Delta T}$ values. More specifically, $D=\{d_{30},d_{60},\dots ,d_{1440}\}$⁠.

Finally, we optimize based on two criterion: first of all we want the lowest $C{V}_{\Delta T,p}$ value which indicate a participants’ data in a specific time-window has the least variation, which in terms of a biometric system could demonstrate that the date is systematic and repeatable. For example, considering a 24-h time-window, it could be expected that the participants average values might be similar for each daily cycle, indicating that they undertake repeatable behaviour. This can be acquired through min(CV_p). The second criterion is to select the greatest $d_{\Delta T}$ value which demonstrates a greater deviation between participant datasets for a given time-window. This allows us to identify where there is the most variation between participants’ datasets. We collapse these two values into a single metric, m, as demonstrated in the following equation:

$$m=\text{max}({\text{min}(\text{CV}}_p)-\text{max}(D))$$

(5)

Table 1 provides the top 10 m values from processing participant data. As evident in the table, the time-windows with the greatest variation between datasets is 14 h. For this reason, in this research, we adopt a time-window of 14 h as being the one with the greatest variation between participants and thus the one where differences between participants can be more easily discovered.

Table 1:

Top 10 m values and their associated time-window duration in hours, illustrating that a 14-h time-window is the most suitable choice

${\text{CV}}_{\Delta T,p}$	$d_{\Delta T}$	m	Time-window (h)
0.19	0.72	0.53	14.00
0.21	0.72	0.51	12.00
0.30	0.73	0.43	10.00
0.36	0.78	0.42	8.50
0.47	0.87	0.40	6.00
0.25	0.64	0.39	11.50
0.17	0.55	0.39	14.50
0.25	0.63	0.38	11.00
0.43	0.81	0.38	6.50
0.07	0.42	0.35	18.50

In total there are 48 datasets.

Now that a 14-h time-window is adopted, participant data are split into 14-h time-window segments, which include calculating the mean heart rate, step count, light sleep, deep sleep and awake sleep values. In other words, the 14-h time-window data are collapsed into five individual measures. An excerpt for a 28-day period is presented in Fig. 3 for all five participants. Figure 3a provides the average heart rate values for each 14-h time-window per participant. As evident in the figure, there is a noticeable difference between each participant; however, there is a good degree of repeatability for each participant. Figure 3b illustrates the total number of steps for each participant within each 14-h window. It is evident once again that there is noticeable difference between the participants, apart from the outlier of participant 4 (black line) who recorded an entry over 20 000 in dataset number 33. Three further graphs are displayed detailing the number of light sleep, deep sleep, and minutes awake during sleep for each of the 14-h time-window dataset. From these graphs, it is evident that there is distinguishable difference between the five participants.

Figure 3:

Example values for the five participants, each represented by a different colour, covering a 28-day period when using a 14-h time-window. (A) Heart rate, (B) steps, (C) light sleep, (D) deep sleep and (E) awake sleep

Application of machine learning

Now that a technique has been devised to time-window the data based on the analysis of historical data, we undertake an experimental analysis to establish the potential of using several supervised machine learning algorithms to correctly match 14-h time-window segments to the correct participant.

Method

In this work, a suite of machine learning algorithms from the WEKA framework [40] are utilized to conduct and evaluate user authentication within the acquired biometric data. User authentication is a binary classification problem [41], where a classifier is trained for each participant. The classifier is provided with a set of recordings from the legitimate user, and also from the users that are impostors. Both of these recordings are used for testing whether the given user belongs to the legitimate user-group or not. Furthermore, all recordings are correctly labelled with participant names, and there is an equal or balanced distribution of classes in the training dataset as each participant provided three months of minute-by-minute data. The missing data values are generated using LI and LE as previously described in ‘Heart rate’ section.

The classification algorithms and their taxonomy types are demonstrated in Table 2. Our motivation for performing empirical analysis on different algorithms is to determine, analyse, and evaluate their performance based on certain metrics. Although research studies, such as the one presented in ref. [42], do exist demonstrating the suitability and applicability of several machine learning algorithms on different types of problems (e.g. data types and complexity), we adopt a systematic exploratory approach testing multiple algorithms across different taxonomy types. All classification algorithms from WEKA framework have been selected which are applicable to our biometric dataset. Our process also aligns to common practice in the biometric research field where researchers often explore the use of multiple supervised machine learning algorithms to identify the most suitable, such as the work in performing iris recognition [43].

Table 2:

Accuracy results from performing supervised machine learning with a 14 h time-window

#	Algorithm	Taxonomy type	TPR	FPR	ACC	EER	ROC area	F1 score
1	Random forest [46]	Tree-based	0.791	0.06	0.7914	0.1652	0.953	0.79
2	Classification via regression [47]	Tree-regression based	0.774	0.062	0.7741	0.1964	0.935	0.772
3	LMT [48]	Tree-based	0.774	0.064	0.7741	0.1875	0.932	0.774
4	J48 [49]	Tree-based	0.773	0.065	0.7727	0.1964	0.918	0.772
5	REPTree [50]	Tree-based	0.784	0.066	0.7842	0.1977	0.882	0.783
6	LogitBoost [51]	Tree-regression based	0.753	0.074	0.7525	0.192	0.921	0.753
7	PART [52]	Tree-rules based	0.751	0.074	0.7511	0.2158	0.874	0.751
8	Iterative classifier optimizer [53]	Tree-regression based	0.727	0.086	0.7266	0.2143	0.912	0.726
9	Random tree [54]	Tree-based	0.744	0.075	0.7439	0.1424	0.835	0.745
10	Decision table [55]	Rule-based	0.698	0.08	0.6978	0.2009	0.914	0.698
11	Multi-layer perceptron [56]	Neural network based	0.709	0.077	0.7094	0.2723	0.878	0.711
12	JRip [57]	Rule-based	0.698	0.097	0.6978	0.2324	0.853	0.697
13	Simple logistics [48]	Function-based	0.685	0.105	0.6849	0.2545	0.881	0.684
14	Logistic [58]	Function-based	0.679	0.104	0.6791	0.2545	0.883	0.679
15	Linear-SVM [59]	Function-based	0.669	0.111	0.6691	0.2813	0.885	0.67
16	IBk [60]	Instance-based	0.627	0.1	0.6273	0.2321	0.821	0.629
17	KStar [61]	Instance-based	0.581	0.125	0.5813	0.2589	0.827	0.578
18	Bayes Net [62]	Bayes-based	0.56	0.131	0.5597	0.308	0.84	0.558
19	SMO [63]	Function-based	0.561	0.175	0.5612	0.247	0.773	0.504
20	Naive Bayes [64]	Bayes-based	0.519	0.149	0.5194	0.3304	0.781	0.515

The performance is measured by using the standard cross-validation evaluation technique, which is an organized method of running repeated percentage splits when testing supervised machine learning techniques [44]. The dataset is divided into 10 pieces (or folds). Each piece is used for testing, whereas the algorithm is trained on the remaining nine pieces collectively. This evaluation outputs 10 values, which are averaged to provide an overall result. If the proposed process was to be used in a live biometric authentication system, then the system would extract the previous 14 h of biometric data from that specific instance through the subjects’ smartwatch, consolidates it as described in ‘Time-window selection’ section and then identifies the subject through the trained model. It should be noted here that the proposed biometric solution does not wait for 14 h to perform the authentication. It simply retrieves the existing biometric data from the smartwatch and decides instantly (depending on the computational power) whether the data are from a legitimate/authorized user or an impostor.

To assess the performance of all algorithms used in this empirical analysis, the following six measures [45] are considered: true-positive rate (TPR) or sensitivity; false-positive rate (FPR) or specificity; cACC; EER; receiver operating characteristic (ROC) area; and finally, F-measure or F₁ score.

Results

Table 2 presents the results from the different machine learning techniques. The taxonomy type of each algorithm is also provided for readers interested in understanding patterns in algorithm performance. The algorithms listed in the table have been sorted in best-to-worst order based on higher TPR, ACC, ROC area and F₁ score values, whilst at the same time, lower FPR and EER values. These metrics represent the consolidated values from processing the five participants. The reason behind including these metrics, especially EER and F₁ score, is to fully understand and analyse the system as the potential impact of misclassification in terms of false acceptance and rejection is different. If the biometric system does not allow access to a legitimate user, it would not be a security risk. However, if the biometric system allows access to unauthorized user, then it would become a critical security issue.

Considering the results, it is clear that tree-based algorithms outperform all other categories for this particular type of biometric data. The ‘Random Forest’ algorithm performed the best, whereas ‘Naive Bayes’ performed the worst within the selected algorithms. The Random Forest algorithm is based on a large number of (ideally uncorrelated) decision trees that work as an ensemble. The large groups of trees are collectively more knowledgeable and informed than individual trees in terms of assigning classes and eliciting predictions. The Random Forest ranked ‘steps count’ and ‘awake during sleep’ features to the top importance (50%) level, whereas light sleep, heart rate and deep sleep were given the second most important (47%) level. Furthermore, the algorithm initially estimated 78.99% ACC and 19.96% EER, but after improving the predictions by using out-of-the-bag estimates, the algorithm produced 79.14% ACC and 16.52% EER. Both ACC and EER values indicate that the accuracy of this biometric system is not substantial due to the limitations in data collection and lack of using any feature extraction technique. The performance can be improved by adding new features or creating from the existing ones, such that it introduces more attributes and uniqueness in the biometric data itself.

Summary of Results

Although this experiment was conducted with only five participants, the data were acquired over a period of 3 months. Furthermore, we have tested 20 algorithms from different categories during feasibility analysis. Therefore, we believe that these results can provide useful insight into creating and deploying a real-time biometric security system using fitness and activity tracking data, as well as providing a repeatable benchmark for the scientific community.

The summary of our results and findings are as follows:

1. the 1- min data points are not adequate for discriminating participants. It is necessary to define a large-enough time-window that can represent distinct samples to obtain results, which in this research was a 14-h time-window;

2. as the biometric data can be acquired from relatively cheap hardware, there is a possibility of having inaccurate, noisy and missing data points. Therefore, it is important that both training and testing data are filtered are pre-processed to normalize and insert any missing values before training and matching;

3. the ROC curve area of all algorithms is higher than 77%. This means that there is a 77% chance or above that every trained model will be able to differentiate participants based on their data and

4. for ACC, the true positives/negatives are important, whereas for F₁ score, the false positives/negatives are important. The values of both ACC and F₁ score are quite similar in all algorithms, which demonstrates that the test results are balanced and do not favour true or false predictions.

In addition, the following recommendations as to future areas of focus that will help produce biometric systems based on fitness and activity tracking data sources:

1. the scarcity of features is one of the major cause behind poor performance. In general, there should be a sufficient number of independent and unique set of features that also do not grow disproportionate with the size of training samples. This would result in rather complex trained models and prevent over-fitting related issues;

2. The dataset utilized for training should be from subjects that are diverse in terms of jobs, ROUTINES, habits, etc. This will create a more representative dataset by encompassing a wider set of situations and possibilities. This may help the algorithms identify and learn patterns which are unique to each participant;

3. along with the feature extraction, systematic mechanisms should also be applied to prioritize features. A set of quality features will be better predictors, hence, resulting in improved accuracy and

4. a biometric security system based on fitness and activity tracking data will require retraining as several factors (age, change in human body, environment, weather, etc.) can affect the extracted heart rate, steps, and sleep values.

Conclusion

In conclusion and reverting to our primary research aim, this investigation has demonstrated the potential of using fitness and activity tracking data to identify individuals. Based on the presented approach of collapsing data sampled at a 1 min frequency into 14-h segments (named time-windows), we tested the acquired data on 20 different algorithms with respect to six performance metrics. Random Forest achieved maximum ACC (79.14%), whereas Naive Bayes performed the worst with the minimum accuracy (51.94%). Also significant is that we were able to determine several interesting characteristics of fitness and activity tracking biometrics data along with the suitability of certain types of algorithms. Our results and analysis demonstrate the potential of using supervised machine learning alongside fitness and activity tracking data to identify individuals, which could be used for identification and authentication purposes in a security context.

There are however many avenues of future research to expand this study to overcome limitations. It would be advantageous to collect more data from a wider participant base. In this research, we acquired data from five participants; however, to increase the confidence in the approach, a higher number of participants are required. Furthermore, in this study, we acquired data for a 3-month period, and although this is a significant volume of data compared with previous studies, a larger range is needed to build further confidence in the presented approach. We propose in the future to revisit our technique to determine the optimal time-window with larger participant numbers. It would also be beneficial to consider the use of alternative devices to judge their impact on accuracy, as well as trialling an increasing range of supervised machine learning algorithms to determine if a greater level of accuracy can be achieved.

Footnotes

1

Statistica state 53 million shipped in 2018: https://www.statista.com/topics/4393/fitness-and-activity-tracker/.

2

Data hosted here: URL to be provided in final version.

3

https://www.mi.com/uk/mi-band-3/specs/.