-, On the digital-computer classification of geometric line patterns,



Similar documents
USE OF EDUCATION TECHNOLOGY IN ENGLISH CLASSES

TEMPORAL PATTERN IDENTIFICATION OF TIME SERIES DATA USING PATTERN WAVELETS AND GENETIC ALGORITHMS

SELF-EVALUATION FOR VIDEO TRACKING SYSTEMS

Chapter 8: Regression with Lagged Explanatory Variables

Automatic measurement and detection of GSM interferences

The Application of Multi Shifts and Break Windows in Employees Scheduling

Making a Faster Cryptanalytic Time-Memory Trade-Off

Measuring macroeconomic volatility Applications to export revenue data,

Journal Of Business & Economics Research September 2005 Volume 3, Number 9

PROFIT TEST MODELLING IN LIFE ASSURANCE USING SPREADSHEETS PART ONE

Predicting Stock Market Index Trading Signals Using Neural Networks

Real-time Particle Filters

GoRA. For more information on genetics and on Rheumatoid Arthritis: Genetics of Rheumatoid Arthritis. Published work referred to in the results:

This is the author s version of a work that was submitted/accepted for publication in the following source:

Using of Hand Geometry in Biometric Security Systems

Statistical Analysis with Little s Law. Supplementary Material: More on the Call Center Data. by Song-Hee Kim and Ward Whitt

DDoS Attacks Detection Model and its Application

Principal components of stock market dynamics. Methodology and applications in brief (to be updated ) Andrei Bouzaev, bouzaev@ya.

Distributing Human Resources among Software Development Projects 1

Segment and combine approach for non-parametric time-series classification

Duration and Convexity ( ) 20 = Bond B has a maturity of 5 years and also has a required rate of return of 10%. Its price is $613.

Individual Health Insurance April 30, 2008 Pages

Market Liquidity and the Impacts of the Computerized Trading System: Evidence from the Stock Exchange of Thailand

Chapter 1.6 Financial Management

Morningstar Investor Return

Single-machine Scheduling with Periodic Maintenance and both Preemptive and. Non-preemptive jobs in Remanufacturing System 1

SPEC model selection algorithm for ARCH models: an options pricing evaluation framework

Performance Center Overview. Performance Center Overview 1

Towards Intrusion Detection in Wireless Sensor Networks

ANALYSIS AND COMPARISONS OF SOME SOLUTION CONCEPTS FOR STOCHASTIC PROGRAMMING PROBLEMS

User Identity Verification via Mouse Dynamics

The Transport Equation

Analogue and Digital Signal Processing. First Term Third Year CS Engineering By Dr Mukhtiar Ali Unar

Forecasting, Ordering and Stock- Holding for Erratic Demand

Course Outline. Course Coordinator: Dr. Tanu Sharma Assistant Professor Dept. of humanities and Social Sciences

A Natural Feature-Based 3D Object Tracking Method for Wearable Augmented Reality

Internet Engineering. Jacek Mazurkiewicz, PhD Softcomputing. Part 1: Introduction, Elementary ANNs

Vector Autoregressions (VARs): Operational Perspectives

TSG-RAN Working Group 1 (Radio Layer 1) meeting #3 Nynashamn, Sweden 22 nd 26 th March 1999

Term Structure of Prices of Asian Options

Diane K. Michelson, SAS Institute Inc, Cary, NC Annie Dudley Zangi, SAS Institute Inc, Cary, NC

Mathematics in Pharmacokinetics What and Why (A second attempt to make it clearer)

Task is a schedulable entity, i.e., a thread

DOES TRADING VOLUME INFLUENCE GARCH EFFECTS? SOME EVIDENCE FROM THE GREEK MARKET WITH SPECIAL REFERENCE TO BANKING SECTOR

Multiprocessor Systems-on-Chips

Efficient One-time Signature Schemes for Stream Authentication *

Unsupervised approach to color video thresholding

The naive method discussed in Lecture 1 uses the most recent observations to forecast future values. That is, Y ˆ t + 1

Making Use of Gate Charge Information in MOSFET and IGBT Data Sheets

Idealistic characteristics of Islamic Azad University masters - Islamshahr Branch from Students Perspective

Identify and ranking the factors that influence establishment of total quality management system in Payame Noor University of Lordegan

Time Series Analysis Using SAS R Part I The Augmented Dickey-Fuller (ADF) Test

Acceleration Lab Teacher s Guide

Multi-camera scheduling for video production

Model-Based Monitoring in Large-Scale Distributed Systems

Chapter 2 Problems. 3600s = 25m / s d = s t = 25m / s 0.5s = 12.5m. Δx = x(4) x(0) =12m 0m =12m

Impact of scripless trading on business practices of Sub-brokers.

Hedging with Forwards and Futures

A Note on Using the Svensson procedure to estimate the risk free rate in corporate valuation

STUDY ON THE GRAVIMETRIC MEASUREMENT OF THE SWELLING BEHAVIORS OF POLYMER FILMS

Child Protective Services. A Guide To Investigative Procedures

ARCH Proceedings

Why Did the Demand for Cash Decrease Recently in Korea?

Supplementary Appendix for Depression Babies: Do Macroeconomic Experiences Affect Risk-Taking?

The Greek financial crisis: growing imbalances and sovereign spreads. Heather D. Gibson, Stephan G. Hall and George S. Tavlas

Appendix A: Area. 1 Find the radius of a circle that has circumference 12 inches.

A New Type of Combination Forecasting Method Based on PLS

DYNAMIC MODELS FOR VALUATION OF WRONGFUL DEATH PAYMENTS

How To Predict A Person'S Behavior

DYNAMIC ECONOMETRIC MODELS Vol. 7 Nicolaus Copernicus University Toruń Ryszard Doman Adam Mickiewicz University in Poznań

Module 3 Design for Strength. Version 2 ME, IIT Kharagpur

4 Convolution. Recommended Problems. x2[n] 1 2[n]

Strategic Optimization of a Transportation Distribution Network

Hotel Room Demand Forecasting via Observed Reservation Information

Molding. Injection. Design. GE Plastics. GE Engineering Thermoplastics DESIGN GUIDE

Trends in TCP/IP Retransmissions and Resets

Chapter 4: Exponential and Logarithmic Functions

Forecasting Model for Crude Oil Price Using Artificial Neural Networks and Commodity Futures Prices

CALCULATION OF OMX TALLINN

WATER MIST FIRE PROTECTION RELIABILITY ANALYSIS

An Agent-based Bayesian Forecasting Model for Enhanced Network Security

Premium Income of Indian Life Insurance Industry

Double Entry System of Accounting

SURVEYING THE RELATIONSHIP BETWEEN STOCK MARKET MAKER AND LIQUIDITY IN TEHRAN STOCK EXCHANGE COMPANIES

Option Put-Call Parity Relations When the Underlying Security Pays Dividends

INTRODUCTION TO FORECASTING

Usefulness of the Forward Curve in Forecasting Oil Prices

Transcription:

IEEE TRANSACTIONS ON PP;ITERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 12, NO. 12, DECEMBER 1990 1217 M. S. El-Wakil and A. A. Shoukry, On-line recogniion of handwrien isolaed Arabic characers, Paern Recogniion, vol. 22, 1989. J. Feder, Languages of encoded line-paerns, Inform. Conr., vol. 13, 1968. T. J. Founain, Array archiecures for iconic and symbolic image processing, in Proc. 8h In. Conf Paern Recogniion, 1986. H. Freeman, On encoding of arbirary geomeric configuraions, IEEE Trans. Elecron. Compu., vol. EC-IO, 1961. -, On he digial-compuer classificaion of geomeric line paerns, in Proc. Na. Elecron. Conf, 1962. K. S. Fu, Synacic Mehods in Paern Recogniion. New York: Academic, 1974. -, Synacic Paern Recogniion, Applicaions. New York Springer-Verlag, 1977. -, Synacic Paern Recogniion and Applicaions. Englewood Cliffs, NJ: Prenice-Hall, 1982. K. Fukunaga, Inroducion o Saisical Paern Recogniion. New York: Academic, 1972. J. S. Huang and K. Chuang, Heurisic approach o handwrien numeral recogniion, Paern Recogniion, vol. 19, 1986. P.J. Knobe and R.G. Wiley, A linguisic approach o mechanical paern recogniion, in Proc. IEEE Compu. Conf, 1967. L. Lam and C. Y. Suen, Srucural classificaion and relaxaion maching of oally unconsrained handwrien ZIP-code numbers, Paern Recogniion, vol. 21, 1988. R. Manara and L. Sringa, The EMMA sysem: An indusrial experience on a muliprocessor, in Languages and Archiecures for Image Processing, M. G. B. Duff and S. Levialdi, Eds., New York Academic, 1981. J. Manas, An overview of characer recogniion mehodologies, Paern Recogniion, vol. 19, 1986. N. J. Nillson, Learning Machine-Foundaions of Trainable Paern- Classifying Sysem. New York McGraw-Hill, 1965. J. F. O Callaghan, Problems in on-line characer recogniion, in Picure Language Machines, S. Kaneff, Ed. New York: Academic, 1970. T. Pavlidis, Analysis of se paerns, Paern Recogniion, vol. 1, 1968. -, Synacic feaure exracion for shape recogniion, in Proc. 3rd In. Join Conf Paern Recogniion, 1976., Srucural Paern Recogniion. New York Springer-Verlag, 1977. S. R. Ramesh, A generalized characer recogniion algorihm: A graphical approach, Paern Recogniion, vol. 22, 1989. H. Same, The quadree and relaed hierarchical daa srucures, Compu. Surveys, vol. 16, 1984. G. S. Sebesyen, Decision Process in Paern Recogniion. New York: Academic, 1968. M. Shridhar and A. Badreldin, Recogniion of isolaed and simply conneced handwrien numerals, Paern Recogniion, vol. 19, 1986. L. Sringa, LCD: A formal language for consrain-free hand-prined characer recogniion, in Proc. 4h In. Join Conf Paern Recogniion, 1978., Efficien classificaion of oally unconsrained handwrien numerals wih a rainable mulilayer nework, Paern Recogniion Le., vol. 10, 1989., A srucural approach o auomaic primiive exracion in handprined characer recogniion, in Proc. In. Workshop Froniers in Handwriing Recogniion, Monreal P. Q., Canada, 1990. L. Sringa and C. Furlanello, A rainable mulilayer nework wih high compuing efficiency, IRST Tech. Rep., 1988. C. Y. Suen, Disincive feaures in auomaic recogniion of handprined characers, Signal Processing, vol. 4, 1982. -, Characer recogniion by compuer and applicaions, in Hand- book of Paern Recogniion and Image Processing. New York: Academic, 1986. C. Y. Suen, R. Shinghal, and C. C. Kwan, Dispersion facor: A quaniaive measuremen of he qualiy of handprined characers, in Proc. In. Conf. Cybern. and Soc., 1977. S. L. Xie and M. Suk, On machine recogniion of hand-prined Chinese characers by feaure relaxaion, Paern Recogniion, vol. 21, 1988. E. F. Yhap and E. C. Greanias, An on-line Chinese characer recogniion sysem, IBM J. Res. Develop., vol. 25, 1981. T. Y. Young and T. W. Calver, Classificaion, Esimaion and Paern Recogniion. Amserdam: Academic Elsevier, 1973. Compuer-Access Securiy Sysems Using Keysroke Dynamics Saleh Bleha, Charles Slivinsky, and Bassam Hussien Absfruc-This correspondence describes a new approach o securing access o compuer sysems. By performing real-ime measuremens of he ime duraions beween he keysrokes when a password is enered and using paern recogniion algorihms, hree on-line recogniion sysems were devised and esed. WO ypes of passwords were considelwl: phrases and individual names. A fixed phrase was used in he idenificaion sysem. Individual names were used as a password in he verificaion sysem and in he overall recogniion sysem. All hree sysems were esed and evaluaed. The idenificaion sysem used en voluneers and gave an indecision error of 1.2%. The verificaion sysem used 26 voluneers and gave an emr of 8.1% in rejecing valid users, and an error of 2.8% in acceping invalid users. The overall recogniion sysem used 32 voluneers, and gave an error of 3.1% in rejecing valid users and an error of 0.5% in acceping invalid users. Index Terms-Compuer-access, securiy, sysems. I. INTRODUCTION The dependence on compuers o sore and process informaion makes he ask of securing access o compuer sysems of grea imporance. Hence, effecive, low-cos, and easily inegraed mehods of assuring valid access have many imporan applicaions [l]. Based on he assumpion ha people ype in uniquely characerisic manners, he presen sudy invesigaes a new approach for building on-line sofware recogniion sysems. Alhough handwriing and yping are differen and are performed in differen environmens, boh are characerisic of he individual performing hese asks. Research on handwriing recogniion sysems idenified basic echniques used for he presen sudy [l], [ 111 - [ 161. Effecive echniques using correlaion analysis of he wriing-hand movemens, pressure measuremens on he signaure surface, and power specrum esimaion of wrien scrips are reaed in he lieraure. The same research shows, however, ha he use of paern recogniion algorihms does no give encouraging resuls when applied o wrien scrips. Neverheless, he use of paern recogniion mehods were found o be effecive in he recogniion sysems presened here. A block diagram showing he general approach used in his sudy is given in Fig. 1. To access a compuer sysem he user is asked o ener a password. In his sudy we reaed wo ypes of passwords: names and fixed phrases. By performing real-ime measuremens of he ime duraions beween he password characers enered we firs obain he measuremens vecor. The feaure-exracion process is hen applied o he measuremens vecor o obain he feaures of ineres in he measured daa. Following he feaure-exracion, a classificaion scheme is incorporaed using he feaures jus obained o perform comparisons wih a se of sored reference paerns. Nex, he decision logic uses he resuls of he classificaion and a predeermined hreshold value o decide on acceping or rejecing he enered password. The hreshold value depends on he classifier used and he recogniion sysem under consideraion. A general recogniion sysem can include boh idenificaion and verificaion. Based on he measuremens jus obained, he idenificaion operaion ries o idenify he user by finding a mach beween Manuscrip received Ocober 5, 1988; revised March 1, 1990. Recommended for accepance by R. De Mon. The auhors are wih he Deparmen of Elecrical and Compuer Engineering, Universiy of Missouri-Columbia, Columbia, MO 65211. IEEE Log Number 9036963. 0162-8828/90/1200-1217S01.OO 0 1990 IEEE

1218 IEEE TRANSACTIONS ON PA ITERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 12, NO. 12, DECEMBER 1990 Reference files + Password Feaure exracor I Classificaion a Decision logic Class labeling Fig. 1. General recogniion sysem. hose measuremens and sored daa for all users. The verificaion operaion, on he oher hand, performs he compuaions required in comparison wih one reference file only; his single reference file is se up for he ideniy claimed in he enered password; he user is hen acceped or rejeced based on a predeermined hreshold value. The idenificaion analysis considers parameers such as he password lengh, he dimensionaliy-reducion or feaure-exracion procedure, he number of users in he idenificaion sysem, and he ype of password which would provide a minimum misclassificaion error. Two classifiers are used, namely he minimum disance classifier and he Bayes classifier. The verificaion analysis focused on searching for a hreshold value for each classifier which, when used, would provide he means o disinguish beween valid users and invalid users known o he verificaion sysem. The experimenaion proocol used in his research included wha messages he subjecs were asked o ype (individual names and a fixed phrase), in wha environmen, and over wha period of ime. Considerable aenion was given o providing informaion and moivaion for he paricipans o produce he quaniy of daa required for all sysems devised and esed. The sudy was pariioned ino wo phases. Phase One was comprised of an idenificaion and verificaion analysis, and provided he guidance o parameerize and es hree online recogniion sysems, which were buil in Phase Two. In Phase Two an updaing procedure was included in all hree sysems considered, and he paricipans used hese sysems ineracively. he personal compuer was used boh o obain he measuremens vecor, and o perform all oher compuaions required for he hree recogniion sysems. The programs were wrien primarily in Forran; in addiion, 8086/8088 microprocessor assembly language subprograms were wrien. These subprograms make use of he sofware keyboard inerrup, and provide he calling programs (Forran programs) wih he ime duraions beween keysrokes. For example, if he password enered is he word HELP he assembly language program compues he acual ime inervals aken beween enering he characers H & E, E & L, and L & P. To have an accepable level of performance for experimens of his kind, i is necessary o ake ino accoun elemens such as he general sae or condiion of he subjecs, he specific aciviy, and he conex in which he aciviy is performed [2]. The compuer was reserved exclusively for his sudy in a separae office. We ried o moivae he voluneers o paricipae by using free coffee and chocolae bars. This helped dramaically in having hem paricipae more readily. An open period of ime was allowed for individuals o paricipae, depending on heir availabiliy and willingness o conduc he experimen. This randomizaion of he order of he experimenaion ends o average ou he effec of he uncorrelaed sources of noise inroduced by he subjec or he insrumenaion [3]. The proocol also ends o average ou he effec of oher sources of noise such as minor illness, faigue, ec. For he hree sysems of Phase Two he reference files were updaed weekly using he laes enries in order o keep up wih he subjecs variaions in performance over ime. In Phase One, he daa was colleced from nine voluneers over a period of nine weeks. In Phase Two, en voluneers esed he idenificaion sysem over a period of five weeks, and weny-six voluneers esed he verificaion sysem over a period of eigh weeks. For he overall recogniion sysem of Phase Two, daa colleced from en valid users during he period of esing he verificaion sysem was used for evaluaing he sysem and 22 voluneers esed he sysem ineracively as invalid users. 111. CLASSIFICATION ALGORITHMS In he decision-making process in paern recogniion, he Bayes classifier, which is opimal in he sense ha is use gives he lowes probabiliy of commiing a classificaion error [4], is frequenly used. The Bayes classifier assumes ha he paern vecor describes a mulivariae probabiliy densiy funcion in he n-dimensional feaure space (n is he dimension of he paern vecor). Because of he difficuly involved in esimaing he densiy funcions from a limied amoun of raining daa, a parameric form of he probabiliy densiy, such as Gaussian, is ofen assumed [5]. The Gaussian probabiliy densiy of a paern vecor X for he user i is given by 11. EXPERIMENTATION PROTOCOL Typing proficiency was no a requiremen in his sudy, alhough each paricipan had a leas an elemenary familiariy wih yping. The sudy was explained o all voluneers, and hey were insruced o be consisen in enering he required informaion. The message o be enered (he password) was no visible during he acion of yping; hence i was necessary o display he enered message on he CRT afer he paricipan enered i. The paricipan was asked o cancel he enry if i was yped incorrecly, and incorrec enries were no processed. One of he prime goals was o choose a password ha is easy o ype and simple o memorize. Such a password can be yped wih a minimum amoun of pauses and/or hesiaions. Paricipans names were chosen as one of he passwords since his choice has been proved o be effecive in handwriing verificaion sysems [l]. Also, a fixed phrase (UNIVERSITY OF MISSOURI COLUMBIA) was used. The experimens were carried ou using an IBM Personal Compuer. In Phase One, he personal compuer was used only o collec he daa (ime duraions beween keysrokes). The daa colleced was ransferred o a VAX compuer for analysis. In Phase Two, so ha each p,(x) is compleely specified by is mean vecor m, and is covariance marix C, for each user i. The mean vecor and he covariance marix for user i are esimaed by: N% m, = (I/N*) x3 3=1 N, C, = (l/n,) x,,xf, - m,ml,=l where N, is he number of paerns used o compue he esimae of he mean vecor and he covariance marix. Assuming recogniion among equally likely users, Bayes decision rule specifies choosing he user whose densiy was mos likely o have generaed he es vecor X [5], [lo]. For recogniion, user i is seleced if p.(x) > p,(x) for all users, i # 3. (1)

IEEE TRANSACTIONS ON PA'ITERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 12, NO. 12, DECEMBER 1990 1219 By aking he naural logarihm of boh sides of (1) and eliminaing like erms, he decision rule is furher simplified and rewrien as follows. Decide user i if d;(x) < dj(x) where di(x) is given by for all users, i # j d,(~) = (X- m,)~x-'(~ - m,) +In IC%/. Pracical experience in speech recogniion sysems indicaes ha he erm InlC,I does no provide significan informaion [6], [7]. This is also rue in his sudy [19]. The decision rule is hen furher simplified o d,(x) = (X - m,yc,-'(x - m,) < d3(x) for all users, i # j. (3) If in addiion C is he ideniy marix (which is no he case here), (3) is reduced o he decision rule of he minimum disance classifier [4], and, for recogniion, user i is seleced if D;(X) = (X- m;)f(x - m,) < D3(X) for all users, i # j. (4) For a verificaion sysem a hreshold value is needed. This value depends on he similariy measure being used. Hence, we mus find ha hreshold value which provides a separaion beween he rue measure and he false measure. The measure is rue when X is provided by a valid user and he reference paerns are compued using his reference file, and he measure is false when X is provided by an invalid user claiming he ideniy of a valid user. Threshold values are hus needed in deermining he rue rejecion error, Type I, when a valid user is rejeced, and he false accepance error, Type 11, when an invalid user is acceped. Iv. PARAMETERIZATION This secion summarizes Phase One resuls for boh he idenificaion analysis and he verificaion analysis [ 191. For idenificaion, he following aspecs were examined: he password lengh, he amoun of daa needed for he reference files o generae he reference paerns, he effec of he ype of password, and he reducion of he dimensionaliy of he paerns. All hese parameers were sudied using boh classifiers, namely he Bayes classifier (3) and he minimum disance classifier (4). The resuls showed ha he longer he password, he smaller he misclassificaion error for boh classifiers. As shown in Fig. 2, he Bayes classifier shows an advanage over he minimum disance classifier for a shorer password. A very long password, however, may be oo long for he users o ype wihou random pauses and/or hesiaions. The misclassificaion error decreases as he number of enries used o esimae he reference paerns increases. This is because of correspondingly beer esimaes for he mean vecor and he covariance marix. On he average, beer performance was achieved when he idenificaion analysis considered individuals' names, raher han he single chosen phrase, as he password. These resuls may have occurred because each voluneer yped a unique password wih which he or she is familiar. I is common in paern recogniion o consider feaure exracion echniques before designing an effecive paern recognizer [4], [lo], [17]. Fisher's linear discriminan and he Karhunen-Loeve ransform (K-L) are usually used for his purpose. In our case he assumpion ha all paern populaion share idenical means precluded he use of K-L [4]. A misclassificaion error of 1.1% was achieved by using Fisher's linear discriminan o reduce he dimensionaliy of he paerns followed by he minimum disance classifier. This echnique 1 w 5 10 Password lengh Fig. 2. Percenage misclassificaion error. (a) The normalized minimum disance classifier. (b) The normalized Bayes classifier. required a longer password han provided by he use of individual names. The fixed phrase was used and he dimensionaliy of he paern vecors was reduced from 30 o 8 [19]. For he verificaion analysis, in addiion o he parameers discussed hus far, a hreshold value is needed. Boh ypes of verificaion errors, Type I and Type 11, are a funcion of his hreshold value. Figs. 3 and 4 show his dependency on he hreshold value. As he Type I1 error decreases, he Type I error increases. I should be noed ha a normalizaion procedure was inroduced for boh classifiers used; ha is o say, he disance measured is divided by he produc of he norm of he paern vecor X and he norm of he mean vecor m. This normalizaion procedure is simply o allow users o choose a password of a variable lengh. The effec of password lengh was also examined. For a fixed value of he hreshold and a fixed number of enries used o compue he reference paerns, he Type I1 error decreased significanly wih he password lengh while he Type I error seemed o be sable. Moreover, for a fixed lengh of he password and fixed hreshold value wih a varying number of enries o generae he reference paerns, he Type I1 error seemed o be sable while he Type I error decreased considerably. We also examined he use of a fixed phrase for he verificaion analysis and found ha he Type I1 error was higher compared o using names as passwords. v. ON-LINE RECOGNITION SYSTEMS This secion presens he resuls of he on-line experimenaion of Phase Two. A close examinaion of he acual disribuion of he measured daa of Phase One showed ha he bulk of he daa has a disribuion wih a Gaussian shape wih ouliers owards he high measuremen values [19]. These ouliers are caused by noisy measuremens due o random pauses and/or hesiaions during he acion of enering he password. We would be more effecive in designing he idenificaion sysem if we could idenify where in he enered password hese noisy measuremens occurred [9]. On he oher hand hese idiosyncrasies are desirable for he verificaion sysem when invalid users are performing, since invalid users end o have more pauses and hesiaions when hey ype ohers' names. Hence, he use of wo enries was incorporaed in he implemenaion of he following sysems. A. The Idenificaion Sysem In he idenificaion sysem wo separae enries were aken from each user for each aemp o be idenified. The wo enries were

1220 IEEE TRANSACTIONS ON PATERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 12, NO. 12, DECEMBER 1990 where Here, X is he es vecor conaining he oucome of combining he wo enries, mi is he mean vecor for user i, and T is a predeermined hreshold value. A hreshold value of 0.08 was found o be saisfacory. User i is idenified as he user who has provided he paern X if (5) is saisfied for all i no equal o j, and he disance measured is less han T. The fixed phrase was employed, and only he firs 15 measuremens were considered for classificaion. Five enries were used o generae he mean vecor for each user. The sysem was esed by en users. The reference files were updaed weekly using he laes en enries. An error was considered o have occurred when he sysem was unable o classify he aemp made as one of he en users because i exceeded he hreshold value. Table I summarizes he resuls of he idenificaion sysem. 0.06 0.07 0.08 0.09 Threshold Fig. 3. Percenage verificaion errors using he normalized minimum disance classifier. (a) Type I error. (b) Type I1 error. b B. The Verificaion Sysem For he verificaion sysem all users were given he opporuniy o choose heir own passwords, and all agreed on choosing heir names; he names used varied in lengh from 11 o 17 characers. Each user was given wo rials for each aemp o be verified by he verificaion sysem. The normalizaion procedure suggesed in (6) and (7) below accommodaed he variaion in he name lenghs, and a differen hreshold value was used for each of he wo classifiers. The normalized minimum disance classifier used a hreshold value of 0.030 on he firs rial, which was hen reduced o 0.029 on he second rial. The normalized Bayes classifier used a hreshold value of O.ooOo30 on he firs rial and a value of 0.000029 on he second rial. Each user file consiss of 30 enries o generae he mean vecor and he covariance marix. The reference daa for each user was updaed weekly using he laes 30 enries o compue he reference paerns. The normalized minimum disance classifier used for verificaion is The normalized Bayes classifier was also used, in he form -- +---- b /? b -5 4 5 6 7 x 10 Threshold Fig. 4. Percenage verificaion errors using he normalized Bayes classifier. (a) Type I error. (b) Type I1 error. hen combined ino a single enry by choosing he shorer of he wo corresponding ime inervals beween keysrokes. This echnique is referred o as he shuffling procedure. Even hough i may seem plausible o average hese wo enries, averaging showed less favorable resuls [19]. The normalized minimum disance classifier given by (5) below was used for he online idenificaion sysem. User i is idenified if D;(X) < Dj(X) < T for all users, i # j Here, he paricipan is claiming o be user i, X is he es vecor, mi is he compued esimae of he mean vecor, C; is he compued esimae of he covariance marix, and Ti and TZ are he hreshold values. Based on he ideniy claimed, he verificaion sysem uses one reference file o compue he esimaes of he mean vecor, he covariance marix, and is inverse; i hen compues he disances using he es vecor and he reference paerns. Boh he normalized Bayes classifier and he normalized minimum disance classifier are hen used o decide on acceping or rejecing he enry. The enry is rejeced if he hreshold values are exceeded for boh of he classifiers used. The sysem was esed by 14 valid users, and a oal of 25 invalid users. Thireen of he invalid users were familiar wih he sysem and were allowed o pracice imiaing he person whose ideniy hey were claiming. The resuls of esing he verificaion sysem are given in Table 11. The able shows boh ypes of verificaion errors namely, he rejecion of valid users ('Qpe I) and he accepance of invalid users (Type 11). The effeciveness of he sysem varied from user o user. For example, wo valid users were no familiar wih yping on a personal compuer keyboard; he only ime hey used his keyboard was when hey came o paricipae in he experimens. These users produced a large Type I error. Their performance improved in he las four weeks

IEEE TRANSACTIONS ON PATIERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 12, NO. 12, DECEMBER 1990 1221 TABLE I IDENTIFICATION SYSTEM RESULTS Aemps Toal Number Aemps Where of Aemps Correcly No Decision Made Classified Was Made % Error 171 169 2 1.17% compared o he firs four weeks of heir paricipaion. In fac, 23 ou of he 44 rue aemps rejeced were caused by hese wo paricipans. Also, one user who had no yping skills (a slow, wo-finger ypis) was easily imiaed by ohers. Fifeen ou of he 22 false aemps were acceped on his name. The proraced absence of paricipans (for a period of en days in some cases), background noise (such as oher people alking while a subjec was in he process of yping he password), and unfamiliariy wih he keyboard all enhanced he Type I error. On he oher hand, he fac ha he paricipans were secrearies and graduae sudens who were known o each oher and were highly moivaed by he gaming aspec of his sudy and aemped o imiae one anoher s ideniies caused he Tjpe I1 error o be high. C. The Overall Recogniion Sysem The overall recogniion sysem is based on he observaions gahered during he period of esing he performance of he wo sysems presened above. Firs, he sysem parameers were se and is performance was deermined by using daa colleced from en valid users during he period of esing he verificaion sysem. Then, 22 voluneers used he sysem ineracively as invalid users by yping he names of he en valid users. For each aemp o be recognized, wo enries were considered. As i appears from he block diagram of he overall recogniion sysem (Fig. 5), here are four sages of rejecing a user. Sage One rejecion deals wih yping errors and uninended enries; ha is o say, his rejecion is an indicaion ha he name enered is no one ha he sysem is designed o recognize. Sage No rejecion uses he same echnique as he on-line idenificaion sysem; here, if a user is no idenified as one of he valid users, hen he or she is rejeced. This is done by having a specified hreshold value of 0.09 and using he normalized minimum disance classifier given by (5). Sage Three rejecion occurs when a user maches a valid user oher han he one being claimed; Sage Three rejecion is hus a misclassificaion error. If a user passes he hree previous rejecion sages, hen we apply he verificaion analysis of he verificaion sysem discussed previously by using (6) and (7). Here, in Sage Four, he normalized minimum disance classifier uses a hreshold value of 0.035 and he normalized Bayes classifier uses a hreshold value of 0.000035. These hreshold values are 10% higher han hose used by he previous verificaion sysem. Table I11 summarizes he resuls of esing he sysem. Wih regard o valid users, seven of he welve valid aemps rejeced were caused by he same user who had experienced major psychological sress during he period of esing he verificaion sysem. Also, for he invalid users, a considerable number of he aemps (70%) were rejeced in Sage No where hey were unable o be mached wih any of he en users known o he recogniion sysem. The oal number of aemps provided by he invalid users was 288, wih 68 aemps (24%) being wrong enries rejeced by Sage One of he recogniion sysem. In he able hese 68 aemps were no considered. VI. CONCLUSIONS In his sudy, we addressed wo major quesions. Firs, how can we characerize he differences in he way people ype on a compuer keyboard, in a manner suiable for idenificaion and verificaion? Sec- Rejecion of Valid Users (Type I Error) Toal aemps = 539 TABLE I1 VERIFICATION SYSTEM RESULTS V Yes Accepance of Invalid Users (Type I1 Error) Toal aemps = 768 Aemps rejeced = 44 Aemps acceped = 22 % Error = 8.1% Sage 1 Sage S rage Sage 4 START (Password) - I Form Measuremen Vecor I I + % Error = 2.8% Valid User Yes Yes Verificaion Analysis 1 Yes Access Graned I END Fig. 5. The overall recogniion sysem block diagram. TABLE I11 OVERALL RECOGNITTON SYSTEM RESULTS Toal Number of Rejeced Acceped Users Aemps Aemps Aemps % Error Valid 380 12 368 3.1% Invalid 220 219 1 0.5% ond, how can we design and parameerize appropriae and effecive recogniion sysems? For idenificaion he choice of he phrase used was user-dependen and all users agreed on i because of heir familiariy wih i. By using one enry per user, he dimensionaliy reducion echniques applying Fisher s linear discriminan followed by he minimum disance classifier provided good resuls (1.1% of misclassificaion error). Wihou dimensionaliy reducion, he same resuls were achieved by requiring he users o ype he password wo imes.

1222 IEEE TRANSACTIONS ON PAmRN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 12, NO. 12, DECEMBER 1990 In he laer case we inroduced a new shuffling echnique, which when applied o boh enries, served o reduce measuremen noise. This shuffling echnique proved o be superior o averaging he wo enries. The online idenificaion sysem gave an unable-o-decide error of 1.17%. For verificaion, he choice of individual names as passwords proved o be effecive. This fac is also rue in handwriing verificaion sysems. Despie he fac ha mos of he paricipans knew each ohers names well, invalid users seemed o have more problems yping ohers names wihou yping misakes. Such yping difficulies held even for users who had good yping skills, possibly because hey are rained on yping words ha are no proper names. Also, long and uncommon names seemed o be very difficul for he invalid users o ype wihou pauses and/or hesiaions. The online verificaion sysem gave an error of 8.1% in rejecing valid users and an error of 2.8% in acceping invalid users. The 2.8% error is relaively high because we allowed he invalid users o wach and pracice imiaing valid users; also, hey were more successful imiaing one paricular user who had no yping skills. Anoher facor enhancing he rejecion of valid users is he fac ha we did no include any raining period for he valid users. Based on he observaions gahered during he period of esing he online idenificaion sysem and he online verificaion sysems, an overall recogniion sysem was implemened and esed. The overall recogniion sysem gave an error of 3.1% in rejecing valid users and an error of 0.5% in acceping invalid users. The implemenaion of all hese sysems required no special hardware and could be accomplished using all ypes of compuers. From he insrumenaion poin of view, his makes he approach compeiive o handwriing recogniion sysems. Recen research on he analysis of wrien scrip shows an error of abou 2% in incorrec verificaion considering six wriers, and oher research shows an error of abou 10% in incorrec verificaion considering foureen wriers [20] - [24]. In his research we used crisp logic. Recenly we have applied he concep of fuzzy logic for classificaion [25]. The resuls achieved are no compeiive o our approach here. REFERENCES N. M. Herbs and C. N. Liu, Auomaic signaure verificaion based on acceleromery, IBM J. Res. Develop., pp. 245-253, May 1977. R. W. Bailey, Human Performance Engineering. Englewood Cliffs, NJ: Prenice-Hall, 1982. C. Hicks, Fundamenal Conceps in he Design of Experimens. New York: CBS College, 1982. J. T. Tou and R. C. Gonzalez, Paern Recogniion Principles. Reading, MA: Addison-Wesley, 1981. D. O Shaughnessy, Speaker recogniion, IEEE ASSP Mag., vol. 3, no. 4, pp. 4-17, Oc. 1986. L. R. Rabiner and R. W. Schafer, Digial Processing of Speech Signals. Englewood Cliffs, NJ: Prenice-Hall, 1978. T. Terzopoulos, CO-occurrence analysis of speech waveforms, IEEE Trans. Acous., Speech, Signal Processing, vol. ASSP-33, no. 1, pp. 5-30, Feb. 1985. J.T. Tou, D. Lu, and J. McCarhy, Auomaic deecion of documen falsificaion, in Proc. IEEE Conk Image Processing and Paern Recogniion, Las Vegas, NV, 1982. D. Hoaglin, F. Moseller, and J. Tukey, Undersanding Robus and Exploraory Daa Analysis. New York: Wiley, 1983. R. Duda and P. Har, Paern Classificaion and Scene Analysis. New York: Wiley, 1973. K. Zimmermann and J. Werner, SIRSYS-A program faciliy for handwriing signaure analysis, in Carnahan Conk Proc., Lexingon, KY, May 1978, pp. 153-155. G. Bills and K. Zimmermann, Specral analysis of righ-handed versus lef handed on-line scrip, in Proc. IEEE Conk Froniers of Engineering and Compuing on Healh Care, Chicago, IL, Sep. 1985. M. Eden, Handwriing and paern recogniion, IEEE Trans. Inform. Theory, vol. IT-8, no. 2, pp. 160-166, Feb. 1962. J. S. Lew, Opimal acceleromeer layous for daa recovery in signaure verificaion, IBM J. Res. Develop., vol. 24, no. 4, pp. 496-511, July 1980. C. Liu, N. Herbs, and J. Anhony, Auomaic signaure verificaion: sysem descripion and field es resuls, IEEE Trans. Sys., Man, Cybern., vol. SMC-9, no. 1, Jan. 1979. J. Verdenbreg and W. Koser, Analysis and synhesis of Handwriing, Philips Tech. Rev., vol. 32, no. 314, pp. 73-78, 1971. K. Han, R. Mchen, and G. Lodwick, The applicaion of an imagecompressiodfeaure ransgeneraion echnique o he compuer-aided diagnosis of brain umors, IEEE Trans. Sys., Man, Cybern., vol. SMC-3, no. 4, pp. 410-415, July 1973. R. McLaren, G. Lodwick, and K. Han, A echnique for feaure ransgeneraion applied o he auomaed diagnosis of medical daa, in Biomedical Symp. Conk Proc., San Diego, CA, 1972. S. Bleha, Recogniion sysems based on keysroke dynamics, Ph.D. disseraion, Univ. Missouri-Columbia, May 1988. H. Almuallim and S. Yamaguchi, A mehod of recogniion of arabic cursive handwriing, IEEE Trans. Paern Anal. Machine Inell., vol. PAMI-9, no. 5, Sep. 1987. H. S. Baird, Feaure idenificaion for hybrid srucural/saisicaal paern classificaion, in Compuer Vision and Paern Recogniion Conk Proc., Miami Beach, FL, June 22-26, 1986. F. Lamarche and R. Plamondon, Segmenaion and feaure exracion of handwrien signaure paerns, in Proc. Sevenh In. Conk Paern Recogniion, Monreal, P.Q., Canada, July 30-Aug. 2, 1984. C. Tapper, Adapive on-line handwriing recogniion, in Proc. Sevenh In. Conf Paern Recogniion, Monreal, P.Q., Canada, July 30- Aug. 2, 1984. G. Loree, On-line handwrien signaure recogniion based on daa analysis and clusering, in Proc. Sevenh In. Conk Paern Recogniion, Monreal, P.Q., Canada, July 30-Aug. 2, 1984. B. Hussien, R. McLaren, and S. Bleha, An applicaion of fuzzy algorihms in a compuer access securiy sysem, Paern Recog. Le., vol. 9, no. 1, pp. 39-43, Jan. 1989.