User authentication on wearable devices by component analysis of heartbeat signals

Today wearable devices, such as smartwatches, become significant part of our life. In spite of rich functionality, computational and device geometry constraints limit the possibility of their usage for strong user-friendly authentication, such as fingerprint or face recognition. Thus, the physiology and behavior based solutions have taken a special interest in last years.

 

Emergence of high-precision sensors for electrocardiogram signals tracking on smartwatches made heartbeat authentication technology an attractive candidate for on-device user authentication. The companies provided advanced solutions for user authentication by heartbeat signals are presented below.

 

 

Nevertheless, proposed solutions for heartbeat signals based user authentication is focused on a single usage context, that may be inappropriate for mass-market products. Thus a suitable alternative is needed for on-device sensitive information processing in any usage context.

 

We proposed to use an innovative component analysis methods to address these issues by applying less resource-intensive algorithms. The considered methods include multifractal analysis (MFA), singular spectrum analysis (SSA), and dictionary learning methods. They provide compact (learned) representation of heterogeneous signals at low computation complexity of signal decomposition procedure.

 

The MFA is aimed at evaluating statistical parameters of heterogeneous signals (multifractals) that provide local power law dependency of adjacency elements values with singularity exponent h(t). The MFA allows for analysing the fine structure of signal by its decomposition at different time scale. The example of obtained generalized fractal dimensions spectrum Dq for different users and various usage contexts are presented below.

 

 

 

In contrast to MFA, Singular Spectrum Analysis (SSA) method is aimed at decomposing of a signal (time series) into a sum of series, so that each component can be identified as either a trend, periodic or quasi-periodic component, or a noise. Analysis of extracted components allows for selecting appropriate components to be used as features during user authentication. Example of estimated SSA spectrum for heartbeat signals gathered for a single user in various usage contexts are shown below.

 

 

Nevertheless, these components should be manually selected, which may be inappropriate in real applications. Therefore it can be useful for decomposition methods that “learn” appropriate representation from the signal. In our research, we used dictionary-learning methods for creating user-specific basis for HBS decomposition. The main idea is that heartbeat signals of target user may be sparsely represented in learned dictionary, while signals of another person require much ”denser” representation. The examples of learned atoms for a single user are presented below.

 

Considered methods were tested on Samsung Galaxy Watch 3 smartwatch. Evaluation was done for two cases: heartbeat signals (HBS) captured in fixed (supine) position, and signals collected in various usage contexts. The first was used to compare the accuracy against the existing state-of-the-art. The second one allows for evaluating performance metrics during real-life activities such as sitting, still standing and walking.

 

Obtained multifractal spectrum, SSA spectrum and decomposition coefficients in learned dictionary were used as features for HBS classification during user authentication. Proposed solution was compared with state-of-the-art Discrete Wavelet Transform (DWT) based method proposed by Enamamu et al. The Enamamu’s solution uses statistical parameters of DWT coefficients for ECG-signals as features during user authentication. Heartbeat signals features were extracted using well-known wavelet bases, such as Daubechies, biorthogonal, Meyer to name a few.

 

The examples of HBS for fixed (supine) position were taken from publicly available datasets found in PhysioNet service:

• Combined measurements of ECG, Breathing and Seismocardiograms – the dataset includes measurements of 20 presumably healthy volunteers. Approximately, 867 signals per-user were captured.
• Wilson Central Terminal ECG Database – the dataset includes ECG signals recorded from 92 patients (14 signal-per-user on average) with I – III Leads.
• The in-house dataset - includes ECG-signals captured with Samsung Galaxy Watch 3 by presumably healthy volunteers (25 logs-per-user with 30 heartbeat-per-log on average) in widespread usage contexts (during lying, sitting, still standing and walking).

 

The estimated FAR and FRR values for considered signal transformation methods and state-of-the-art Enamamu’s solutions are presented below.

Dataset	Used features	FAR, %	FRR, %
Wilson Central Terminal ECG Database	Multifractal analysis	4.97	8.16
	Singular spectrum analysis	0.08	11.95
	Dictionary learning	70.95	6.33
	Enamamu et al. solution	3.61	2.08
Combined measurements of ECG, Breathing and Seismo- cardiograms	Multifractal analysis	0.03	1.77
	Singular spectrum analysis	0.01	0.68
	Dictionary learning	0.42	51.62
	Enamamu et al. solution	0.01	0.77

 

Note large values of FAR (FAR ~ 0.5%) and FRR (FRR~ 50%) metrics for Dictionary Learning method in comparison with other methods. This can be explained by high similarity of captured HBS that leads to high similarity of learned dictionaries A for users. On the other hand, MFA and SSA methods allow for achieving authentication accuracy on open datasets that is close to the state-of-the-art Enamamu’s solution.

 

Also, False Acceptance Rate and False Rejection Rate were calculated by using HBS captured for various usage contexts (user’s physical activities). Evaluated metrics for considered signal processing methods and Enamamu’s solution are presented in the next table.

 

Dataset	Used features	FAR	FRR
Users are lying	Multifractal analysis	2.66	39.90
	Singular spectrum analysis	5.49	28.06
	Dictionary learning	0.21	60.95
	Enamamu et al. solution	1.76	22.15
Users are sitting	Multifractal analysis	2.14	30.28
	Singular spectrum analysis	4.12	20.53
	Dictionary learning	2.68	58.62
	Enamamu et al. solution	0.95	18.79
Users are standing	Multifractal analysis	6.54	29.51
	Singular spectrum analysis	3.61	17.95
	Dictionary learning	0.56	63.06
	Enamamu et al. solution	0.88	14.25
Users are waling	Multifractal analysis	34.69	20.50
	Singular spectrum analysis	7.70	36.98
	Dictionary learning	3.41	76.66
	Enamamu et al. solution	2.03	45.37

 

Using Dictionary Learning techniques leads to considerable increase of FRR values (up to 76%) by preserving FAR values close to state-of-the-art solutions. Usage of multifractal and singular spectrum analyses allowed to achieve close to Enamamu’s method authentication accuracy.

 

Evaluation results showed that proposed processing pipelines, based on multifractal analysis and singular spectrum analysis, achieved close to state-of-the-art Enamamu et al. solution accuracy (0.01% False Acceptance Rate and 0.77% False Rejection Rate), and outperformed it in various usage contexts. Obtained results confirm the effectiveness of using specialized signal processing methods for heartbeat authentication on constrained devices such as wearables.