Cardiac Cycle Biometrics using Photoplethysmography

Cardiac Cycle Biometrics using Photoplethysmography Emiel Steerneman University of Twente P.O. Box 217, 7500AE Enschede The Netherlands e.h.steerneman@student.utwente.nl ABSTRACT A multitude of biometric verification methods exist, based on fingerprints, faces, or other biometric features. These methods can be spoofed with a piece of clay or a photo. This research will explore the viability of measuring a heartbeat using photoplethysmography (PPG) on fingers for biometric verification. A pulse oximeter was used to measure the amount of blood in fingers of individuals under equal conditions, over an interval of ten seconds. For preprocessing, the bias was removed and the measurements were normalized. Using a data set of 60 measurements, and cosine similarity as a classifier, an Equal Error Rate of 22.3% was achieved. The results indicate that verification using these methods cannot meet the performance of current biometric systems. It might however, depending on the application, complement existing systems by introducing a liveness check and an extra level of verification. Keywords Biometrics, Photoplethysmography (PPG), Verification 1. INTRODUCTION Biometrics is the authentication of humans based on their unique biological characteristics. These characteristics can be found all over and in the body. A few example characteristics are fingerprints, DNA, veins, and voice [1]. Biometrics is nothing new [7]. There is evidence that around 500 B.C., fingerprints were used by the Babylonians and the Chinese to identify people. Egyptians used physical descriptors to identify traders. Around 1800, the Bertillon system was implemented, in which a physician measured various body dimensions and stored these on easily sortable cards. Over the last few decades, use of dependable Biometrics systems has been on the rise due to the technological advances of computing. Fingerprint scanners can often be found on laptops and phones, iris scanners are employed by airports, and many devices support voice and facial recognition [4, 5]. Biometric identification methods are generally safer than conventional methods such as passwords and passphrases, since these do not rely on a piece Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. 26 th Twente Student Conference on IT February 3 rd, 2017, Enschede, The Netherlands. Copyright 2017, University of Twente, Faculty of Electrical Engineering, Mathematics and Computer Science. of data that can easily be lost or stolen. Spoofing Still, biometric identification methods are not foolproof [8]. Fingerprints can be faked by something as simple as a piece of clay. Face and eye recognition can be spoofed by using printed pictures. Voices can be recorded. The ability to spoof existing biometric identification methods has led to the search for more robust unique biological characteristics. One of these characteristics might be the heartbeat. Heartbeats have the advantage that they can be measured in many different ways. Also, in contrast to a fingerprint or a piece of clay, the person has to be alive. More important, heartbeats are harder to record unknowingly, and thus harder to duplicate. A well-known measuring method which has been used for years, in many different fields and applications, is the Electrocardiogram (ECG). An ECG is a record of the heart s electrical activity, and can be measured by placing electrodes on different parts of the body. This technique is already used in wearables such as bracelets to measure heart rate. However, getting a reliable reading can be quite cumbersome and intrusive. Other less popular methods to measure heartbeats are Seismocardiograms (SCG) which measure the vibrations of the heart, and Phonocardiograms (PCG) which record the sounds that the heart makes. PPG Another method that is gaining attention is Photoplethysmography (PPG). PPG uses infrared light to measure the amount of blood flowing through a certain area. More blood absorbs more of the emitted infrared light, which is measured by an infrared sensor. Success has been achieved by measuring infrared light shined directly through the heart. However, this method is invasive, can be uncomfortable, and the long term health risks are not clear. Another option is to measure the heartbeat in a finger. When placed on a finger, the sensor measures the amount of blood currently present in the finger. Since the amount of blood in the finger is influenced by the heartbeats, the data reflects these heartbeats. this data can be used for biometric verification. In comparison to other methods, PPG sensors can be relatively small and cheap, and can be manufactured for handheld devices such as smartphones. 1.1 Related work PPG as a verification method does not reach the results achieved by methods that are currently used, such as verification based on fingerprint or face. Gu and Zhang reached a false reject rate of 17% [6]. Another team of researchers managed to achieve an Equal Error Rate of 5.29% [2]. Capelli et al., 2010 [3] rated different fingerprint-based verification systems, and multiple systems manage to reach an Equal Error Rate of 3% or less. The gap in this performance can be partially explained by the emotional state 1

having an effect on the heartbeat morphology. Research has been done on this problem, and showed promising results [9]. 1.2 Problem Statement Until the performance of PPG-based biometrics is on par with fingerprint-based biometrics, it is not effective to use it to secure critical applications. It is possible that PPGbased biometrics can accompany fingerprint-based biometrics, thereby enhancing security with an extra layer. PPGbased biometrics add a liveness check to the process, together with making sure that the vital signals are from the same individual that places a finger on a fingerprint scanner. A good application for this would be the unlocking of smartphones. Many smartphones already come with functionality to measure the heartbeat and a fingerprint scanner, and software could simply combine the two. Fingerprint scanners are hardware peripherals with their own processes, whereas analyzing the heartbeat would have to be done on the processor of the mobile phone (or any other handheld device or wearable) itself. To ensure that the performance load on processors is limited, the biometric performance of cosine similarity as a relatively simple and light classification algorithm was analyzed. Cosine similarity also does not need a large data set, which was limited by time constraints. 1.3 Research Question Based on the problem statement, the research addresses the following research questions: 1. What is the Equal Error Rate of cosine similarity as a classifier, given the measurements made using PPG? 2. How does the Equal Error Rate of cosine similarity as a classifier compared to stae-of-the-art biometric systems 3. Does the heart rate have an effect on the morphology of the heartbeat? 2. METHODS 2.1 Data collection To test the possibility of verifying individuals based on the heartbeat in their finger, it is best to have measurements with the least amount of noise as possible. The clearest measurements will give the best performance, and conclusions can be drawn with the assumption that the system will always perform equally or worse. Measurements were taken of six individuals, where each individual made 10 measurements for a total of 60 measurements. each individual was measured within the same hour and placed under the same conditions. Sensor The sensor used was a pulse oximeter. A pulse oximeter can be placed on the finger, and uses infrared light to measure the level of oxygen saturation, which is proportional to the amount of blood. The sensor measured at a rate of 12KHz. The duration of each measurement is 10 seconds, resulting in 12 x 10 4 samples per measurement. The sensor is supposed to be connected to a smartphone using a 3.5mm audio jack, but in this research it is connected to an external audio card, which is connected to a laptop. Because the sensor works with an audio jack, it was made to support frequencies that humans can hear. These frequencies range between around 100 Hz and 20KHz. All other frequencies are filtered out by low-, and high pass filters in the audio card. Since the base frequency of a heart lies around 1 Hz, it is filtered out and cannot be measured. The solution was to modulate the heartbeat onto a carrier signal of 12 KHz which passes through the filters, and then use demodulation to separate the heartbeat from the carrier signal. Unfortunately, modulation and demodulation introduced noise, and possibly removed features that make the heartbeat unique. For unknown reasons, the measurements are upside down, when compared to how saturation is normally represented. A lower value in the measurements means a higher level of saturation. This does not affect the biometric performance as long as all measurements are upside down. Influence of sensor noise The sensor introduced significant noise in the measurements. Measurements that would certainly throw off the classification had to be removed. These measurements had peaks in them that prevented the preprocessing step from correctly identifying individual heart beats. The origin of these peaks is unknown. A possibility is faulty wiring where movement makes wires touch, leading to a short circuit. Another issue with the sensor was that the first few moments of a measurement would consist only of very strong noise. This problem was solved by simply removing the first two seconds of each measurement. Another challenge introduced by the sensor was that individual heartbeats would not be on a horizontal line. This can be seen in figure 6. The amplitude and offset of the signal also fluctuated between different measurements, and sometimes also within the same measurement. Influence of heart rate One measurement was done after an individual ran two laps around the house. When measuring, the individual had a heart rate of 150 bpm. The corresponding signal was clearly different from the 60 other signals that were measured at around 60 bpm. The signal looked like it was flattened out. Out of all the signals, this signal came the closest to looking like a single period of a sine wave. This can be see in figure 4. 2.2 Preprocessing In the preprocessing phase, a raw measurement like the one in figure 1 is converted into a single, averaged, normalized heartbeat. The preprocessing phase consists of multiple steps, which are described below. 2.2.1 Detecting heartbeats The first step of preprocessing is detecting heartbeats in a measurements A heartbeat can be detected by finding the two steep descents that mark its start and end. These are found by taking a small interval of a measurement, and calculating its slope. If this slope passed a certain threshold, the x-coordinate of the center of the interval is marked as the start of a new heartbeat. The interval consists out of 500 samples, with a duration of around four milliseconds. The interval was varied between 100 and 2000 samples, and 500 samples was found to work for every measurement made. This interval might be too long for higher heart rates. Because the heart rate is higher, the heartbeats are faster and the interval of heartbeats are shorter. Therefore, the interval of the downwards slope indicating the start of a heartbeat will be shorter. Thus, the interval would also have to be shorter. Picking a single slope threshold which would work for any measurement is not possible. As stated in Data collection, 2

the amplitudes of different measurements are not consistent. This causes different heartbeats to have different slopes at their start. If the threshold is too high, the starting point of a heartbeat might not be detected. If the threshold is too low, points will be detected that are not starting points of a heartbeat. As can be seen in figure 2, a heartbeat consists of two bumps, a small one and a larger one. The descent of the small bump is especially prone to being detected if the threshold is too low. The problem is solved by picking a high threshold, and gradually lowering it until three conditions are met. The factors by which the threshold is lowered when a given condition is met were chosen experimentally, and worked well for all the measurements made. Condition one The first condition is that there has to be more than one starting point in the measurement. The duration of a heartbeat is obtained by calculating the distance between the starting point of this heartbeat and the next. If there is only one starting point, then the heartbeat cannot be measured. In this case, the threshold is lowered by a factor of. Condition two The second condition is that the heartbeats share roughly the same interval. For example, if one heartbeat has an interval that was twice the length of the other heartbeats, then this is an indication that this heartbeat is actually two heartbeats, and that these two are not detected because of the threshold being too high. In this case, the threshold is slightly lowered, by a factor of 0.95, because it is already very close to detecting all the heartbeats correctly. Condition three The third condition is that there is not an abnormally high or low beats per minute (bpm) in the measurement. Each measurement used in this research was done while the individuals were at rest, which means that the heart rates of the measurements should lie somewhere within the range of 50 to 80 bpm. Everything outside of this range is considered to be an abnormal bpm. If the bpm is too low, this means that two or more heartbeats were identified as one. This has to occur multiple times in equal intervals for it to pass the second condition, and is therefore highly unlikely. It has not happened in any of the measurements. If the bpm is too high, than this means that each heartbeat is detected as two or more heartbeats. This is reasonable when the threshold is too low. As stated before, a heartbeat has two bumps, where the first small bump can be falsely identified as a starting point of a heartbeat, turning an actual heartbeat into two. This was observed multiple times when the threshold was still fixed. Figure 1: Raw measurement Figure 3: Heartbeats without the vertical bias Figure 4: Blue heartbeats at 60 bmp, Red heartbeat at 150 bpm 2.2.2 Removing bias and reducing noise The second step is to remove the bias and noise present in each heartbeat. Removing heartbeat bias The first bias is a vertical offset, which can be seen in figure 2. Looking at all measurements, it seems like there is some sort of low frequency sine wave present. This wave might be an artefact of the modulation and demodulation described in Data collection. The bias is measured simply by drawing a line between the start-, and endpoint of a heartbeat. From each heartbeat, its corresponding bias is subtracted. The entire measurement is then moved, such that the first sample of the measurement has the position x=0, y=0. The result can be seen in figure 3. Reducing noise as stated in section Data collection, the measurements are quite noisy and have more samples than the classifier needs. Also, the biometric performance of the classifier would be severely impacted with this many samples, because small peaks caused by noise could be seen as a unique features. To reduce the noise and the amount of samples, each detected heartbeat is divided into 50 slices, and the samples in a slice are averaged into a single value. The amount of slices was chosen experimentally. With less than 50 slices, the two bumps present in a heartbeat got flattened too much. With more than 50 samples, noise was clearly visible in the measurement. 2.2.3 Normalizing heartbeats Each heartbeat is normalized so that the results of the classification algorithm can easily be understood and processed. Details and justification can be found in the section Classification. Normalization is done by dividing each of the 50 samples by the Root-Sum-Square value. This value is given by the formula n i=1 y2 i where y i is the value of the i th sample and n is the number of samples, 50 in this case. Each heartbeat is then stretched over the same interval and moved, such that the first sample of the heartbeat has the position x=0, y=0. The result can be seen in figure 5. The stretching of the heartbeats over the interval is irrelevant for the classification algorithm, but it allows for clear visualization of the difference between each heartbeat. Figure 2: Raw measurement, with the detected heartbeats and bias Figure 5: Heartbeats after slicing and normalizing Creating the final heartbeat The normalized heartbeats are added together and aver- 3

aged to create a single heartbeat. This heartbeat is once again normalized. This heartbeat is stored and ready to be used by the classification algorithm. Thirty of these final heartbeats can be seen in figure 6. False Accept Rate Figure 6: 30 heartbeats of three individuals. Each individual has a different color. The heartbeats show a distinguishable difference False Reject Rate Figure 8: False Acceptance Rate against the False Rejection Rate 3. CLASSIFICATION Cosine Similarity was chosen to calculate the similarity between two measurements. This method was chosen because it is relatively low-cost to calculate. Also, it does not need to be trained, which is preferable with a small data set. Another advantage is that a measurement is simply a 50-dimensional vector with values between -1 and 1. Comparing two measurements with cosine similarity gives a value in the range [-1, 1], where 1 is returned when the measurement is compared with itself, and where -1 is returned when the signal is compared with the exact opposite of itself. Since the data is normalized, the cosine similarity function turns into a simple dot product of the two measurements. This is the reason that each measurement was normalized using the Root-Sum-Square value, as stated in the section Preprocessing. The data set of 60 measurements was compared with itself, and a the False Acceptance Rates and False Rejection Rates were measured with different thresholds. The thresholds range from 0.5 to 1, with incremental steps of 01. 4. RESULTS The results of the classification with a data set of 6 individuals with 10 measurements each can be seen in figure 7, 8, and 9. Rate FAR FRR EER 0.5 0.7 0.9 Threshold Figure 7: False Acceptance Rate and False Rejection Rate against the Threshold. The Equal Error Rate lies at 0.945, with rates of 22.3% True Acceptance Rate False Accept Rate Figure 9: Receiver Operating Characteristic (ROC) 5. DISCUSSION Influence of the heart rate Figure 4 shows that the heart rate has a significant impact on the morphology of the heartbeat. The research done by Sarkar, Abbott, and Doerzaph report the same significance [9]. Unique features Looking at figure 6, each individual clearly has unique features in their heartbeats. However, noise from the sensor, modulation and demodulation, and possibly other factors are too much of an influence on the biometric performance of the system. 6. CONCLUSION Performance of the system Concluding from the Equal Error Rate of 22.3%, the system should not be used anywhere on its own. The chance that someone is correctly rejected is 77.7%. Most mobile devices give the user three attempts to unlock it before locking the user out. Given that an intrudes is trying to access the system, the chance that the system rejects the intruder three times is only 46.9%. The false acceptance rate could be lowered by increasing the threshold, but this will make it even harder for genuine users to access the system. Additionally, this research makes the assumption that all measurements are done under the same circumstances. In reality, users will have different heart rates, different emotional states, and place their finger on the sensor inconsis- 4

tently. Performance compared to other systems Compared to other biometric systems, this system performs poorly. One team of researchers managed to reach an Equal Error Rate of 5.29% using PPG-based biometrics [2], and Capelli et al., 2010 [3] indexed fingerprint-based biometric systems with an Equal Error Rate of 3% or less. Further research The system could be used in combination with other biometric systems, such as fingerprint scanners. The other biometric system can verify the user, and the PPG-based system can check for liveness. Possibly, the PPG-based system can also verify if the liveness check is done on the same user that was verified by the other biometric system. This research gives an indication of the performance of a PPG-based biometric system using cosine similarity as a classifier. Due to the small dataset, the results are not conclusive. More research with a larger dataset has to be done to provide more conclusive results. 7. REFERENCES [1] Biometrics Institute Limited. Types of biometrics. http://www.biometricsinstitute.org/pages/ types-of-biometrics.html. [2] A. Bonissi, R. D. Labati, L. Perico, R. Sassi, F. Scotti, and L. Sparagino. A preliminary study on continuous authentication methods for photoplethysmographic biometrics. In 2013 IEEE Workshop on Biometric Measurements and Systems for Security and Medical Applications, pages 28 33, Sept 2013. [3] R. Cappelli, D. Maio, D. Maltoni, J. L. Wayman, and A. K. Jain. Performance evaluation of fingerprint verification systems. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(1):3 18, Jan 2006. [4] B. Clark. The history of biometric security, and how it s being used today. http://www.makeuseof.com/tag/ the-history-of-biometric-security-and-how-its-being-used-today/, March 2015. [5] Find Biometrics. Applications. http://findbiometrics.com/applications/, 2014. [6] Y. Y. Gu and Y. T. Zhang. Photoplethysmographic authentication through fuzzy logic. In IEEE EMBS Asian-Pacific Conference on Biomedical Engineering, 2003., pages 136 137, Oct 2003. [7] S. Mayhew. History of biometrics. http://www.biometricupdate.com/201501/ history-of-biometrics, January 2015. [8] C. Roberts. Biometric attack vectors and defences. Computers & Security, 26(1):14 25, 2007. [9] A. Sarkar, A. L. Abbott, and Z. Doerzaph. Biometric authentication using photoplethysmography signals. In 2016 IEEE 8th International Conference on Biometrics Theory, Applications and Systems (BTAS), pages 1 7, Sept 2016. 5