Authentication Performance Analysis of Acoustic-Based Ear Biometrics System Using Audible Signals

A study of an acoustic-based ear biometrics system for individual identification using audible signals was undertaken. In this era of too much network connectivity, an increase in cybercrimes has led to the loss of key information or leakage of it. Experts have therefore continued to work hard to protect and secure important data from being stolen, lost, leaked, or tempered with. The use of biometric systems is one of the methods that many have adopted to provide the needed security and protection of data. The authors in this study proposed an acoustics-based ear biometric system for the identification of individuals while taking into consideration the dynamics of real-time data capture and using real persons for user authentication applications. The system has been developed in the MATLAB/SIMULINK language which supports dynamic real-time data capture. The results of simulation experiments showed that with proper experimentation and threshold calibration, it is possible to develop acoustics-based authentication systems that can identify individuals correctly and with 100% recognition accuracy. Depending on the human subjects under study, the threshold cosine similarity setting may vary between 0.2 and 0.4. However, this variation is offset by the enrollment procedure deployed in practice. Furthermore, trend analysis using moving average analysis revealed the possibility that a false acceptance is equally likely even though 100% recognition accuracy was attained.

Introduction

In this era of many networks’ connectivity, an increase in cybercrimes has led to loss of key information or leakage of it. Hearable devices like The Dash or Xperia Ear are also rapidly increasing. Experts have therefore continued to work hard to protect and secure important data from being stolen, lost, leaked, or tempered with. The use of biometric authentication is one of the methods that many have adopted to provide the needed security and protection of data [1]. Authentication of users comes in different forms such as voice recognition, use of fingerprint, use of iris or retina, etc., [2], [3]. To capture the user’s identity, each user is expected to expose the relevant body part (the eye, finger, palm, etc.) to a camera or place the palm on a scanner for proper capturing.

Biometrics authentication has been extended to the ear as a part of the human body, and much progress has been made as far as ear biometrics is concerned but different kinds of issues are readily encountered [4]. Besides ear biometrics based on ear recognition, acoustic ear recognition is another powerful or useful authentication method that is based on the exclusive acoustic features of different ear canals [5]. Interestingly, [6]–[9] carried out research that used the acoustic properties of the ear canal and the pinna for the purpose of user recognition.

In this paper, we intend to carry out quiet authentication of users using inaudible sound signals. The reason for the quiet authentication is to avoid interrupting the users. The efficiency of the method will be analyzed. The sections of this work are Section I which is the introduction, Section II which presents the literature review, Section III is the methodology adopted in the work, Section IV is the result presentation, and finally, Section V is the summary and conclusion.

Literature Review

The authors in [1] presented an identification system that used ear acoustic biometrics. The research was aimed at investigating how feasible the use of ear acoustic biometrics that takes advantage of the ear canal’s reflective properties as a Head Related Transfer Function (HRTF) can be when it comes to the recognition of individuals in real-time within dynamic environments. It was shown that with good calibration of threshold and adequate experimentation, the development of an authentication system based on ear acoustics capable of identifying people correctly is possible.

An investigation of fast individual authentication simulator system that considers Ear Canal Impulse Response, known as ECIR for short, and Ear Canal Transfer Function, known as ECTF for short, together with an earbud for the probing of microphone signal was carried out by [6]. The research used the acoustic properties of the ear canal and the pinna for the purpose of user recognition. [10] used certain audio signals that were audible and had a range of frequency of about 1500 Hz for capturing a person’s ear acoustics. It was mentioned that based on the kind of device used for capturing, error rates achieved ranged between 0.8 percent and 18 percent.

Mahto et al. [11] presented a detailed study on ear acoustics biometrics based on in-audible signals. The authors emphasized that even though the user is busy discussing, he can still be authenticated without any negative effect due to background noise. Zhong et al. [12] gave research work on ear acoustics biometric systems for the recognition of individuals through key features of the ear. In the study, an Equal Error Rate of 14.9 percent was obtained at the best performance. Since the error rate was that high, the adopted methods will have few applications in real-world situations.

Materials and Method

Materials

The materials considered necessary for this work include a personal computer, loudspeaker, earphones, sound card, etc.

Method

Based on Fig. 1, there are two signals, the echo signal represented by $y_t$ and the probe signal represented by $x_t$. There is a transmission of $x_t$ through the ear canal while the produced echo signal is captured and recorded.

Fig. 1. Authentication system for the ear [6].

For a pair of echo signals $y_1(t)$ and $y_2(t)$, cross-spectral density CSD from Fig. 1 is used to extract ear acoustic features. Probe signal and echo signals are fully subjected to Fast Fourier Transform (FFT) so as to get the Discreet Fourier Transform of the input signals. These outputs from the FFT are taken to the cross-spectral density where the CSD between the two the inputs $Y_{w}and{X}_w$ is determined to obtain the transfer function of the ear canal. The ear canal transfer function is given by: where $X_{\left(w\right)}^{\ast }$ represents the complex conjugate of $X_{\left(w\right)}$. Based on the output from the CSD which is the ear canal transfer function, Mel-Frequency Cepstral Coefficients (MFCCs) are extracted and subjected to filtering. Then after applying the Discreet Fourier Transform, the Mel-Frequency Cepstral Coefficients are post-processed. The purpose of the post-processing is to get normalized Mel-Frequency Cepstral Coefficients [13]–[15]. Finally, there is a computation of the similarity between ear acoustic features that have been extracted. The equation for cosine similarity is given by:

(1) $$\begin{array}{r}{H}_{w=\frac{Y_{\left(w\right)}{X}_{\left(w\right)}^{\ast }}{X_{\left(w\right)}{X}_{\left(w\right)}^{\ast }}}\end{array}$$

(2) $$\begin{array}{r}{\mathit{C}}_{\mathit{s}}=\frac{C_e^T{C}_t}{\sqrt{(II{C}_{e}II{C}_{t}II)}}\end{array}$$

If ${\mathit{C}}_{\mathit{s}}$ is greater than a certain threshold already fixed, then one person has the two features.

The operation of Fig. 2 is that signals coming from a device like the Loudspeaker positioned near the ear are captured by the ear. The captured signals are pre-amplified by means of a Preamplifier. The sound card parses the pre-amplified signals to a personal computer for signal processing. In addition, the Sound Card provides a probing sound signal which the power amplifier amplifies and sends to the Loudspeaker. Fig. 3 shows a flowchart illustrating the data capture process.

Fig. 2. The architecture of ear biometric system data capture [1].

Fig. 3. Data capture flow chart [1].

Feature Representation Computation and Head-Related Transfer Function

To find the exclusivity of a particular human subject during the analysis of EBS and bearing in mind the modifying effect of the environment on sound pressure waves affecting the human ear, many compute the HRTF, (Head Related Transfer Function) of the considered subjects. With this, it is possible to evaluate the response of the human ear to changes in the excitation of stimulus probes as stated in [12].

(3) $$\begin{array}{r}{H}_L(r,\theta ,\varphi ,f,a)=\frac{P_L(r,\theta ,\varphi ,f,a)}{P_0(r,f)},\end{array}$$

where

(4) $$\begin{array}{r}{H}_R(r,\theta ,\varphi ,f,a)=\frac{P_R(r,\theta ,\varphi ,f,a)}{P_0(r,f)}\end{array}$$

P_L represents left ear sound pressure,

P_R represents right ear sound pressure,

P₀ represents free-field sound pressure,

r represents source distance related to head center,

θ – azimuth,

ϕ represents head elevation relative to center,

f represents source distance related to head center,

a – human subject.

The length of the ear canal, the longest and shortest diameters, and the volume of the ear canal are as shown in Table I.

Table I. Ear Canal Lengths [1]

Parameters	Dimensions
Ear canal’s length	25 (±2.5 mm)
Longest diameter	9.4 (±1.5 mm)
Shortest diameter	4.8 (±0.5 mm)
Ear canal volume	1014.0 (±15.4 mm3)

Procedures for Requirements and Experiment

Preliminary live-test measurements were based on a whole of 4 subjects carried out within a closed room in the Agricultural Engineering Laboratory, Rivers State University (RSU) and Ignatius Ajuru University, all in Rivers State, Nigeria. The test subjects included two females and two males, and some consecutive measurements were taken with each participant reciting the word “Greeting” for 3 different simulation trials.

Earlier there was an implementation live, to test the technique that was developed. In most cases, the proposed protocols’ testing, and evaluation may not be practically achievable by means of real experiments because that would be far more complex, time-consuming, and highly costly. Therefore, to address this problem, “TESTBEDS and SIMULATORs” are effective tools relied upon for testing and analyzing the protocols’ performance [16].

Test Requirements for Dynamic Environment

a)Here, some small rules are made to limit the volume of ear canal changes. These rules are No gum-chewing, minimize laughing, yawning, smiling, etc. b)Avoiding abrupt head movements to keep the headphones intact.

Test Requirements for Static Environment

a)No movement of subjects for at least 10 seconds. b)Any obstruction to ear measurements must be removed.

Gathering of Dataset

To authenticate the concepts proposed in this research, the HRTF dataset obtained from the IRCAM laboratories was utilized. The dataset is composed of a set compensated Windows WAV file and a mat file which includes the structured data file format for easy analysis and comprehension. There are 187 sample points constituting azimuth, elevation, and signal content data captured at a sampling frequency of 44.1 kHz.

Results and Discussion

In an ideal situation, it should be expected that the moving average trend should move downwards from points 4 to 5 as seen in both Figs. 4 and 5, depicting a low or reducing false acceptance rate because of comparison to other subjects different from the base subject. As seen in the trend plots of Fig. 4, it is obvious that there is a somewhat useful indication in subject 1 of the likelihood that a deviation from the expected values is minimal as the cosine values decrease from the 4^th and 5^th comparison samples implying the less likelihood of a false acceptance. However, this is not the case when considering subject 2 as there is a tendency towards false acceptance from the 4^th and 5^th comparison sample trials as seen in Fig. 5.

Fig. 4. Subject 1 trend analysis with the moving average filter.

Fig. 5. Subject 2 trend analysis with the moving average filter.

Table II shows that the cosine similarity values of comparison between counts 2 and 1 are higher than the one of count 3. This can be explained because comparison count 3 is an intersubject recognition of the ear test. Therefore, as we can expect, it should give very low cos values.

Table II. Acoustics-Based Ear Recognition Test–Subject 2 [1]

Comparison count	Cos (a, b)
1	0.2893
2	0.2255
3	0.0573

Based on Table III, counts 1 and 2 comparison represents the acoustics-based intra-subject ear biometric recognition cases for subject 1. It should be noted that the cosine similarity values of comparison between counts 2 and 1 are higher than the ones of count 3. Again, this can be explained because comparison count 3 is an inter-subject recognition of the ear test. As a result, it should give very low cos values.

Table III. Acoustics-Based Ear Recognition Test–Subject 1 [1]

Comparison count	Cos (a, b)
1	0.3782
2	0.4820
3	0.0573

Conclusion

The authors in this study proposed an acoustics-based ear biometric system for identifying individuals while considering the dynamics of real-time data capture and using real persons for user authentication applications. The system has been developed in the MATLAB/SIMULINK language which supports dynamic real-time data capture.

Extensive experiments considering the Head Related Transfer Functions (HTRF) of several human subjects in the Rivers State University and the Ignatius Ajuru University campus environs have been conducted. In particular, the aspect of enrolling person data and considering comparative inter/intra-subject analysis were performed on subjects.

The results of simulation experiments showed that with proper experimentation and threshold calibration, it is possible to develop acoustics-based authentication systems that can identify individuals correctly and with 100% recognition accuracy. Depending on the human subjects under study, the threshold cosine similarity setting may vary between 0.2 and 0.4. However, this variation is offset by the enrollment procedure deployed in practice.

Furthermore, trend analysis using moving average analysis revealed the possibility that a false acceptance is equally likely even though 100% recognition accuracy was attained. Thus, it becomes necessary to conduct extensive simulation experiments to further reveal likely failures in the model of the proposed solution and minimize the problem of over-fitting on datasets.