Biometric Authentication Using Machine Learning Techniques

Abstract

Information technology has drastically improved the information security industry. One way is through biometric authentication which uses physiological and behavioral measures to authenticate users before granting them access to services or systems. This paper presents the various biometric technologies that are based on machine learning techniques. Certain software on the machines detects templates and matches them to existing ones to reject or accept users. It has also outlined some of the biometric authentication technologies whereby further development is ongoing.

Introduction

Authentication is a two-step process whereby the first step may involve typing in a username, plugging a smart card or scanning their fingerprints, palm or iris. The second step is verification whereby a password may be required or reliance on the technology to match the scan to a particular user’s details before allowing access (Fulton, 2011). Authentication uses three types of credential related data. The types that use passwords and pins are problematic because they can be forgotten , lost, copied or stolen The type that uses biometrics or physiological features such as fingerprints, face, iris, voice, palm or vein are user friendly and are hard to steal although they sometimes raise privacy concerns (Fulton, 2011).

Machine learning involves creating systems that are able to learn from data. For instance, such a system may be trained or programmed to classify information according to the features provided as a benchmark identifiers to distinguish between spam mail and ordinary mail. Machine learning methods are used in biometrics applications to construct, recognize and information fusion for purposes of integrating multiple biometrics and obtain rich results and helpful methods for dealing with temporal data (Salah, 2012).

Machine learning focuses on making predictions based on previous behavior or observation samples. Machine learning techniques are engineered to handle complex problems that hand-crafted models are unable to handle. Machine learning provides higher automation in the knowledge engineering process. It is important as its automated detection saves time, is more accurate and efficient as compared to manual human activity. Machine learning involves discovering and exploiting regularities in training data. The most common learning techniques includes numeral networks, case based reasoning, classification and regression trees, rule induction, genetic algorithm and genetic programming. On a general level, two types of machine learning exist, inductive and deductive. Inductive functions by creating computer programs through rules and pattern extraction from massive data sets while deductive functions through the current facts and knowledge and by deducing new information from the old. (Singh, Kumar, & Sangwan, n.d.). Machine learning techniques are used in biometrics by taking the available signals and estimating the probability that they belong to the correct human. The standard RBF learning and algorithms are used to locate RBF networks. This RBF-GA is the hybrid algorithm that is enabled to find function centers outside the input clusters. This indicates that the RBF networks that have the center locations offer more accuracy in recognition than those that have the centers located inside the input clusters (Kung, Mak, & Lin, 2004).

Biometric authentication has become an important security measure because of the increased rate of illegal immigration, border intrusion and fraudulent activity relating to visa among other important security avenues. Biometric authentication technology is increasingly being deployed for security at checkpoints in commercial complex, airports and retail malls. The reason why this form of security is preferred over traditional methods is because it offers accurate verification and allows only a small margin for errors or duplication. The technology is gradually being adopted across all industries with prime security maintenance requirement. Some of the areas are government offices dealing with defense, travel and immigration, and banking and finance institutions. For the government, applications are limited to control building access, verify licenses, personal ID and voting among others. Banks on the other hand use the biometric technology as a control for accessing accounts and Automated Teller Machines among others. The biometric authentication industry is growing and the global total market share is forecasted to grow to $13.89 billion in 2017 (Rohan, 2013)

Companies wishing to restrict unauthorized access to data centers and other related information turn to Biometric authentication technologies. This form of user identification adds an extra layer of security to the traditional keycards and or password system. This form of technology eliminates the need to have security guards and viewers of footage from surveillance cameras. Installation of biometric technologies such as fingerprint scanners in company settings increases the security layer and cuts the cost of having to employ guards. Biometrics technology has become increasingly important even for personalized activities because people are relying heavily on technology to preserve their most confidential information such as financial records. It has thus become necessary for them to boost their security levels beyond mere passwords. For instance, Intel has developed a sensor that scans the vein pattern on people’s palms for use on a laptop or tablet. This technology serves to identify the user before allowing them access and it is a stronger security measure compared to the traditional password (Clark, 2012). This is to make sure that services are only accessible to legitimate users.

Some of the biometric authentication technologies or applications used include finger prints, facial recognition, eye scans and palm screening. There are other technologies that are already in use or at development stages such as the Next Generation Identification (NGI) by the US Federal Bureau for Investigation which is expected to be up and running by 2014. The NGI will use a host of identifiers such as photos that are searchable by use of face recognition technology, iris scans, measures of gait, voice recordings, fingerprints records, scars, tattoos and palm prints. DNA analysis is the other technology currently being explored. (Clark, 2012).

Biometric authentication technologies

Fingerprint technology

This biometric is the most used globally due to the ease in acquisition, unchallenged acceptance and use as compared to other forms of biometrics. There are various forms of fingerprint readers that vary in price and security level. The optical reader is a cheap sensor that is unable to read dirty or marked fingers and it is an easy target for hackers. The Capacitive reader that forms an image of a fingerprint using electrical current and it is harder for hackers to trick it. The Ultrasound reader is by far the most reliable reader as it captures the fingerprints from the dermal skin layer and it does not necessarily require for one to have a clean, unscarred finger and it also very reliable for accuracy reasons. Thermal readers sense the temperature between the finger features such as ridges and valleys. However, all finger print readers are designed to detect the distinct finger print patterns and features. There are three pattern types namely arch, loop and whorl. The features are the important details captured by the readers for recognition ( Biometric Solutions, 2013).

After acquiring the fingerprint image, the matching software interprets the image. The software is designed to efficiently compare and match the read-outs against each other. The software used is of two types. One that matches the specific minutiae features or the one that matches the pattern ( Biometric Solutions, 2013). When identifying a person using this biometric technology the stored template is compared to the provided template for authentication.

Artificial neural networks, support vector machines and genetic algorithms are the contemporary methods used in presenting fingerprint identification solutions. They build a vector that is based on a given rule and this allows the machine to process complicated data from the finger print. Other techniques are the k-Nearest, decision Tree and Linear regression. The performance measures used are the root mean square error, mean absolute error and relative absolute error. It is reported that out of the methods outlined, the decision tree method has the best performance (Molale, twala, & Seeletse, 2013).

Face recognition Technology

This biometric uses a computer application to automatically identify and verify a person using a digital image or a video frame. One of the ways the system works is by locating the position of the facial features such as the eyes, nose and mouth. The application is usually rescaled to a defined size called the canonical image. After capturing the facial measurements, the image is stored in a face template.

The other way is the Eigen method which categorizes faces based on their degree with a fixed set of 100 to 150 Eigen faces. The Eigen faces created appear as dark and light areas arranged in a particular pattern. The pattern indicated how the facial features are different from each other and then it is scored after evaluation the evaluation assesses symmetry, facial hair style, nose or mouth size. These two methods use software that analyzes the face’s spatial geometry. Face authentication is commonly used to identify missing persons and fugitive criminals among others (Bhattachayya, Ranjan, A., & Choi, 2009).

It is notable that RBF networks have more utility in biometrics as compared to LBF networks. For example, the HyperBF network is used in face recognition and it able to report 100 percent recognition rate on a database with over forty people. In contrast, the EBF network performs substantially better on voice recognition or verification as compared to EBF. It is also noteworthy that in terms of superiority in accuracy in identification and training time, RBF is better that LBF. For face recognition a modular network that is based on auto associative RBF is applied and it is based on XM2VTS and ORL face database (Kung, Mak, & Lin, 2004).

Eye scan

The pattern of the veins and arteries in the eye are unique in each person even for identical twins. Near Infra-Red light is used to capture the image of the iris or the retina. The image of the iris is converted into a bit pattern which preserves the required information to compare with a template. This Iris code translates the visible image features into a 512 byte code, which is the template allowing fast searches and low false acceptance rate. For the retina, the vessels are photographed and analyzed in the same way. It is not possible for someone to fake an iris or retina scan which makes this one of the most reliable authentication techniques. The templates or codes are then matched with the template provided for authentication. This is however a relatively new technology that is being refined (Biometrics Solutions, 2013). The machine leaning technique that has made a contribution in the iris biometric is the Multi Vector Machines which uses the Multi Objective Genetic algorithms to select the matching features in the iris accurately. The asymmetrical SVM is used to deal with the false accept ad false reject in a new way as well as balancing the specific classes with respect to the other involved classes (Sheela & Vijaya, 2010).

Speaker recognition

This biometric evaluates the features that produce the voice and not the sound. The features are dependent on the dimensions of the vocal tract, mouth, nasal cavities among other physiological speech determinants. Speaker recognition may use a text dependent or text independent system. The text dependent requires that the person says a fixed phrase which remains constant for enrollment and verification or the system may prompt for a repeat of randomly generated phrase. The text independent one is based of any words that the speaker may choose. The process involves a voice recording, followed by feature or characteristics extraction, then matching the pattern and the decision to accept or reject. The recording may be collected form a dedicated system like a telephone. The system then analyzes the frequency, pitch, duration and volume. During comparison, the extracted frames are matched with the template using pattern matching or a dynamic time warping (Bhattachayya, Ranjan, A., & Choi, 2009). The machine learning algorithms are the MLP,RFBN, C4.5 decision tree and Bayes net. The C4.5 shows consistent performance, the MLP works better for gender recognition and the RFBN offers better performance for large populations’ authentication (Sunil, Shruti, & Krishna, 2010).

Hand geometry technique

This biometric is founded on the knowledge that almost every person has a different hand shape. This technique uses the mechanical optical principle to measure the hand’s length, width, thickness and surface area. The optical scan requires a source of light and a black and white camera to create a 2D bitmap image of the hand and compare the characteristics of the hand. The hand geometry assesses both the vertical and horizontal characteristics of the hand for a 3D image. After this, the image of the hand geometry is then digitalized and matched to a template for authentication (Fulton, 2011).

Palm scan, DNA sampling and Key stroke dynamics

This is similar to an eye scan whereby it obtains the unique vein patterns on a palm and matches or compares them to existing templates. It is currently being developed and when fully developed, it might be a competitor against the eye scan. (Clark, 2012). The veins are able to absorb infrared light and become darker and are displayed as so in the image taken by the infrared camera. DNA sampling is another type that is being explored further. It requires a blood or a bodily sample for biometric analysis. It is however being used in crime detection and law enforcement.

Key stroke dynamics is a behavioral biometric that identifies a person based on the pattern, speed and rhythm of typing on a keyboard. It measures the time one spends while pressing a key and the time spent when releasing and pressing the next key which are technically referred to as dwell time and flight time. The machine learning techniques that depend on support vector machines (SVM) and neural network technology are used to handle keystroke dynamics. It has been noted that the SVM are more accurate than the neural network technique (Zang & Orgun, 2010).

Biometric application is required for authentication and identification. Authentication is whereby verification is required to tell that a person is who he or she may claim to be and for identification to determine who the person may be as measured by the existing biometric data. Biometrics solutions rely on users biological or behavioral measures. They include fingerprints, palm patterns, veins patterns, iris features, and voice and face patterns. Behavioral measures may include mouse or keystrokes dynamics which may analyze aspects such as typing speeds or mouse behavior and this does not require the use of a biometric sensor (Rubens, 2012)

There is a deployment for systems that use sensors on mobile devices such as smartphones and tablets. The sensors used include a camera for face recognition, a microphone which is used for voice recognition and the keyboard for typing rhythm. This system is cost effective because companies are not obligated to purchase biometric hardware. This is because users may already have smartphones and they could use them whenever they require logging in into the company or organization’s system. The Wi-Fi connectivity is commonly used to transmit biometric data on to a back end authentication system (Rubens, 2012)

In as much as use of biometric technology is more secure than traditional authentication methods such as passwords it is never a hundred percent accurate. This is because when users are enrolling, they most often require providing more than one sample such as a fingerprint. This is then used to create a template which is used to authenticate the biometrics provided by the user by assessing whether the sample is similar enough to the template earlier given and stored. The False or Non match and False Match Rate scores are used to measure the accuracy of the system. This shows that the system is likely to give errors and may not be reliable as the only security check in a highly security dependent industry (Rubens, 2012)

Biometric measures are not secrets like passwords are and for this reason, it is possible for hackers to present say a photograph for face recognition biometric and gain access or a wax cast of a fingerprint for a fingerprint scanner or recorded voice for a voice recognition biometric system. New technologies are designed to tackle these risks by having features that may for instance ask one to say random words instead of one specific set of words and this protects against someone else using a recorded file for a voice recognition system. Blinking may be required to ascertain that a face image being used is not a photograph for a face authentication biometric system. Other fingerprint screening biometric system may have features to detect electrical conductivity or measure heat to ascertain that the fingerprints are obtained from a living body and not a wax cast. These additional features are meant to enhance anti-spoofing (Rubens, 2012).

There are rapid innovations in technology that avails better sensors and computer applications for identifying people. The advances serve to address the increasing security needs and they have continued to garner acceptance as privacy and ethical concerns are adequately addressed. The various machine methods applied for the different types of biometrics have been outlined. The methods are used in the various available systems although there is a need to have faster algorithms that are more accurate.

References

Biometric Solutions. (2013). Finger Print Recognition. Retrieved May 14, 2013, from Biometric Solutions: http://www.biometric-solutions.com/solutions/index.php?story=fingerprint_recognition

Bhattachayya, D., Ranjan, R., A., F. A., & Choi, M. (2009, September). Biometric Authetication: A Review. International Journal of u-and e- Service and technology, 2(3), 13-27.

Biometrics Solutions. (2013). Iris Recognition. Retrieved May 14, 2013, from Biometrics Solutions: http://www.biometric-solutions.com/solutions/index.php?story=iris_recognition

Clark, J. (2012, October 2). Latest Biometric Technology. Retrieved May 14, 2013, from The Data Center Journal: http://www.datacenterjournal.com/it/the-latest-in-biometric-technology/

Fulton, J. (2011). Biometric Authentication is Here and Now. Digital Persona, 1-40.

Kung, S., Mak, M., & Lin, S. (2004). Biometric Authentication: A machine Learning Approach. London: Prentice Hall.

Molale, P., twala, B., & Seeletse, s. (2013). Fingerprint Prediction Using Statistical and Machine Learnig. ICIC Express Letters, 311-316.

Rohan, M. (2013). New Report Next Generation Biometric Technologies Market (2012-2017) by Marketsand Markets. PRWEB Online Visibility from Vocus, n.p.

Rubens, P. (2012, August 17). Biometric Authentication: How it Works. Retrieved May 14, 2013, from E-Security Planet Internet Security for IT Pros: http://www.esecurityplanet.com/trends/biometric-authentication-how-it-works.html

Salah, A. A. (2012). Machine Learning for Biometrics. Netherlands: Center for Mathematics and Computer Science.

Sheela, S., & Vijaya, P. (2010). Iris Recognition Methods-survey. International Journal of Computer Science, 0975-8887.

Singh, Y., Kumar, P., & Sangwan, O. (n.d.). A Review of Studies on Machine Learning Techniques. Journal of Computer Science and Security, 1-15.

Sunil, A., Shruti, A., & Krishna, C. R. (2010). Prosodic Feature Based Text Dependent Speaker Recognition Using Machine Learning Algorithms. International Journal of Engineering Science and Technnology, 5150.

Zang, B.-T., & Orgun, M. A. (2010). PRICAI 2010: Trends in Artificial Intelligence. New York: Springer.