Active authentication for mobile devices utilising behaviour profiling

Download Active authentication for mobile devices utilising behaviour profiling

Post on 23-Dec-2016




1 download


  • Int. J. Inf. Secur.DOI 10.1007/s10207-013-0209-6


    Active authentication for mobile devices utilisingbehaviour profiling

    Fudong Li Nathan Clarke Maria Papadaki Paul Dowland

    Springer-Verlag Berlin Heidelberg 2013

    Abstract With nearly 6 billion subscribers around theworld, mobile devices have become an indispensable com-ponent in modern society. The majority of these devices relyupon passwords and personal identification numbers as aform of user authentication, and the weakness of these point-of-entry techniques is widely documented. Active authenti-cation is designed to overcome this problem by utilising bio-metric techniques to continuously assess user identity. Thispaper describes a feasibility study into a behaviour profilingtechnique that utilises historical application usage to verifymobile users in a continuous manner. By utilising a combina-tion of a rule-based classifier, a dynamic profiling techniqueand a smoothing function, the best experimental result fora users overall application usage was an equal error rate of9.8 %. Based upon this result, the paper proceeds to proposea novel behaviour profiling framework that enables a usersidentity to be verified through their application usage in acontinuous and transparent manner. In order to balance thetrade-off between security and usability, the framework isdesigned in a modular way that will not reject user accessbased upon a single application activity but a number of con-secutive abnormal application usages. The proposed frame-work is then evaluated through simulation with results of11.45 and 4.17 % for the false rejection rate and false accep-tance rate, respectively. In comparison with point-of-entry-based approaches, behaviour profiling provides a significant

    F. Li (B) N. Clarke M. Papadaki P. DowlandCentre for Security, Communications and Network Research(CSCAN), Plymouth University, Plymouth PL4 8AA, UKe-mail:;

    N. ClarkeSchool of Computer and Information Science, Edith Cowan University,Perth, WA, Australia

    improvement in both the security afforded to the device anduser convenience.

    Keywords Active authentication Behaviour profiling Biometrics

    1 Introduction

    With nearly 6 billion users globally, mobile devices havebecome ubiquitous in our daily life [17]. The modern mobilehandheld device is capable of providing many multimediaservices through a wide range of applications over multiplenetworks as well as on the handheld device itself. These ser-vices are predominantly driven by data, which is an increas-ingly sensitive nature. Indeed, a series of studies have high-lighted the vulnerability of mobile devices by storing per-sonal information (e.g. date of birth), login credentials (i.e.user name and password) and business data (e.g. corporateintellectual property) [20,23,30]. Therefore, any misuse ofthe services and information stored upon the device poses athreat to its owner when the device is lost or stolen, infectedby malware [26] or attacked by social engineering [13].

    The most popular user authentication approach is the pass-word or personal identification number (PIN). Although thisapproach is available on most mobile devices, a survey con-ducted by [9] demonstrated that 40 % of their participantsfailed to utilise this simple security mechanism. In addition,possible misuse could still occur if mobile users do not utilisethe technique properly, such as never changing the PIN, writ-ing the PIN down or sharing the PIN with others [6,21]. How-ever, the fundamental weakness of the PIN-based approachis that as a point-of-entry technique, it does not validatethe users identity again once the initial trust is obtained.DARPA proposed an Active Authentication program aiming


  • F. Li et al.

    to overcome the problem of the point-of-entry technique byutilising behavioural-based biometrics for desktop comput-ing [10]. According to DARPA proposal, there are a numberof biometrics that have the potential to be applied within themobile device domain, such as gait recognition and keystrokedynamics. One such biometric approach is behavioural pro-filing. Its particular advantage is being independent of theservice being utilised (i.e. you need to be walking with gaitand typing with keystroke analysis). Based upon the findingsfrom a series of feasibility studies, this paper focuses upondemonstrating a novel behavioural profiling framework thatcan provide continuous and transparent protection for mobiledevices.

    This paper begins by introducing the study and presentingthe state of the art in behavioural biometrics that have beenapplied within the mobile domain. The paper then proceeds todescribe a comprehensive experimental study of mobile userapplication usage. Based upon the results, a novel behaviourprofiling framework that will provide the verification of amobile users identity in a continuous and transparent manneris proposed and then evaluated through simulation. The paperconcludes by highlighting the future direction of research.

    2 Behavioural biometrics for mobile devices

    The use of biometrics to authenticate a persons identity onmobile devices has been an established area for more than10 years. In 1998, Siemens and Triodata developed a fin-gerprint recognition prototype with which a user can gaininstant access to the phone by swiping their finger against asensor. In 2011, facial recognition was integrated in the Sam-sung Galaxy Nexus smartphone [25]. However, these phys-iological biometric techniques are mainly deployed to offerpoint-of-entry security. In comparison, behavioural biomet-ric methods have the ability to continuously and transparentlyverify a users identity. However, they also tend to have fea-tures that are noisier, vary more overtime and are subject tovarious environmental aspects that can affect sample collec-tion and classification. So, care must be taken when consider-ing their implementation particularly in active authenticationwhere less control over the user and the environment exists.

    With the evolution of mobile devices, the chance to adoptother biometric-based authentication techniques has becomemore realistic. Indeed, some devices already possess a num-ber of inbuilt sensors that are capable of collecting a vari-ety of user biometric traits, enabling several behaviouralapproaches to be deployed upon them, including behav-iour profiling, gait recognition, handwriting recognition, key-stroke analysis and voice verification. Although research hasbeen carried out on most of the aforementioned behaviouraltechniques [5,7,11,31], little work has been undertaken inthe area of behaviour profiling.

    Behaviour profiling identifies people based upon the wayin which they interact with the services of their mobile device.In a behaviour profiling system, a users current activities(e.g. dialling a telephone number) are compared with anexisting profile (which is obtained from historical usage)by using a classification method (e.g. a Neural Network).The users identity is determined based upon the compari-son result. Therefore, behaviour profiling systems have thepotential to provide continuous and transparent identificationwhile users interact with their device.

    Research into behaviour profiling started around late1990s by monitoring user calling and migration behaviourover service providers networks to detect telephony ser-vice fraud. An overview of studies and their performancecharacteristics are presented in Table 1. For call-based frauddetection systems, a users profile can be created based uponvarious calling features (e.g. International Mobile SubscriberIdentity (IMSI) and start date of call) that have been gath-ered over a service providers network during a period oftime [2,15,24]. The call-based fraud detection system con-stantly monitors each users current calling activities againtheir profile. An alarm will be raised if the deviation betweenthe current calling activity and the users profile exceeds apredefined threshold.

    Based upon the theory that people have a predictablepattern when they travel from one location to another, themobility-based fraud detection system monitors a mobileusers location activities to detect abnormal behaviour (e.g.theft of the device) [3,16,28,29]. A mobility profile can begenerated by using a users historical location informationthat can be obtained either from a mobile cellular networkin the form of cell IDs or via a global positioning system(GPS). Similar to the working principle of the call-basedfraud detection system, an alarm will be raised if the userscurrent mobility activity significantly deviates from his/herprofile.

    By monitoring a users calling or location activities,behaviour-based fraud detection systems offer a high levelof performance in terms of detection rate (DR) and falsealarm rate (FAR) (as demonstrated in Table 1) and detectunforeseen attacks. Also, no additional computational poweris required from the mobile device as the detection proce-dures are carried out within the service providers network.Such solutions have been effective at protecting traditionalmobile devices in terms of fraud as these devices cannotcarry out the detection task by themselves, and they aremainly utilised to access telephony and text message services.As modern mobile devices have the ability to access multi-ple networks and host many applications, the behavioural-based fraud detection systems cannot provide adequate pro-tection for them anymore. Therefore, a new system thatcan offer protection for all networks and applications whichmobile devices connect with is needed. The next section will


  • Active authentication for mobile devices utilising behaviour profiling

    Table 1 Review of networkbased mobile behaviourprofiling

    Pattern classification model DR (%) FAR (%)

    Calls Mathematical formula [24] 80 3Calls Neural networks [15] 50 0.02Calls RBF neural network model [2] 98 4.2Mobility Bayes decision rule [3] 88 Mobility High order Markov model [29] 88 15Mobility Instance based learning [16] 50 50Mobility High order Markov model [28] 89 13

    describe a feasibility study into behaviour profiling withinthe mobile device host environment via user applicationusage.

    3 Behavioural profiling on mobile devices

    It is widely recognised that users utilise mobile services toperform a variety of tasks when interacting with individualapplications. In order to thoroughly investigate the possibilityof employing the behaviour profiling technique within themobile host environment, two types of application behaviourwill be examined for this research:

    Standard applications: provide a basic level of informa-tion on how a mobile user utilises the device, such as thename of an application, the time when the application wasaccessed and the location at which it was utilised.

    Extended applications: offer richer and more discrimina-tory information than standard applications do, i.e. infor-mation regarding what a user does with them (e.g. thetelephone number within a telephone application and aweb address within a web browser application).

    The experiment employed a publicly available dataset pro-vided by the Massachusetts Institute of Technology (MIT)Reality Mining project [12]. The MIT dataset contains arich amount of mobile users application activities over along period of time: 106 participants enroled for the datacollection process from September 2004 to June 2005. Theexperiment utilised a subset of participants whose activi-ties occurred during the period of 24/10/200420/11/2004to maximise the number of participants (i.e. if two users haddifferent start and end dates, the date feature alone would pro-vide the discriminatory information required and skew theexperimental result). These activities include 101 standardapplications and two extended applications [telephony andShort Message Service (SMS)] as shown in Table 2. Despitesome extended applications (e.g. a web browser) beingavailable in the dataset, the extra information (e.g. visited

    Table 2 The MIT datasetStandard apps Telephony SMS

    Participants 76 71 22Unique apps or

    telephone numbers101 2,317 258

    Logs 30,428 13,599 1,381

    website) was not captured; hence, they were treated as stan-dard applications. It was the only open dataset that containedsufficient quantity and quality of data to allow for the experi-ment to be conducted. By utilising the MIT dataset, three setsof experimental studies were conducted: descriptive statisticson mobile applications, a preliminary study of the telephonyservice and behaviour profiling of mobile applications. Thefirst study sought to better understand the nature of the datawith a view to identify possible features for subsequent analy-sis and evaluation. The second study was conducted to iden-tify an optimal classifier for solving the behaviour profilingissue within the mobile host environment and to determinepositive behaviour profiling features. The third study wasdesigned to examine the feasibility of employing behaviourprofiling on mobile devices through application activities.

    3.1 Descriptive statistics analysis

    It has been established that descriptive statistics have theability to determine potential positive features for formingunique patterns to discriminate individual users [18]. In orderto examine potential positive features of both standard andextended applications for behaviour profiling, two sets ofanalysis were conducted on the data presented in Table 2.

    For standard applications, their name, time and locationof access were available from the dataset and examined. Forexample, usage of all users phonebook application featuresis shown in Fig. 1. For the location usage comparison, user71 did not share any cells with any other users (althoughit is difficult to visualise due to the large cell IDs range in arelative small graph). As a result, this user could be identifiedbased upon the location feature alone. On average, each user


  • F. Li et al.

    Fig. 1 Usage comparison of all users phonebook features

    only shared 2 cells with another user. For the worst case,users 41 and 66 shared 10 cells. Although the usage on these10 cells represent 70.8 and 62.5 % of their total usage, theyutilised these cells very differently: user 41 spent the majorityof location usage in cells 1, 77 and 18, while user 66 utilisedcell 49 most of the time. Therefore, the majority of their usagecould still be separated. For the comparison of the time ofaccessing the phonebook application, no clear patterns areshown between individual users in general although the figuredoes highlight usage differences between several groups ofusers (i.e. high, medium and low usages).

    For extended applications, the name of the application,time and location of access, plus extra information the appli-cation offers, were available from the dataset and analysed.For example, Fig. 2 depicts the usage comparison of thelocation and telephone number features of the telephony ser-vice from all 71 users. This feature was comparable withthe phonebook application, with individual users identifiedby utilising the location feature alone. For the comparison ofthe telephone number usage, users 26, 33 and 59 did not shareany telephone number with any other users; also on average,each user only shared one telephone number with anotheruser. Hence, the majority of the users could be discriminatedfrom each other based upon the telephone number featurealone. Some users did share several telephone numbers dur-ing the chosen period (e.g. users 34 and 63 shared 9 telephonenumbers. However, they could still be discriminated user 34heavily used the anonymised telephone number 1237 whileuser 63 employed the anonymised telephone number 1247

    and 1271 most frequently. Regarding the time of call fea-ture, it is difficult to identify individual users by using it inisolation (as with the phonebook application). For the dura-tion of call comparison, no clear discriminative informationis available as the majority of calls lasted

  • Active authentication for mobile devices utilising behaviour profiling

    Fig. 2 Usage comparison on 71 users telephony features

    a call may last (e.g. the relationship between calling partiesand how often they communicate); with the observed per-centages, it is clear that it could be difficult to discriminateusers from each other by using the duration of call alone.A summary of the effectiveness of the aforementioned fea-tures for a potential successful classification of mobile usersis summarised in Table 3.

    3.2 A preliminary study utilising telephony activity

    The selection of effective features is a critical process inpattern classification as the system performance is closelyrelated to the features and a large number of features willincrease the complexity and the size of the classifier resulting

    in a problem known as the curse of dimensionality [1]. Also,the literature on behaviour profiling identified a number ofclassification methods that performed well, and they fall into

    Table 3 Individual application feature contributing towards a success-ful classification

    Features Contributions

    Standard feature Application name PositiveExtended feature Telephone number PositiveStandard feature Location PositiveExtended feature Duration of call NegativeStandard feature Time of call Negative


  • F. Li et al.

    two categories: statistical and neural network approaches.Based upon the no free lunch theorems, there is no singleclassification method that can solve all given problems [31].Three classification methods were chosen from these cate-gories to identify an optimal classifier and explore the effec-tiveness of various application features to solve the behav-iour profiling within the mobile host environment: the radialbasis function (RBF) neural network, the feed-forward multi-layered perceptron (FF MLP) neural network and a rule-based approach.

    As the full dataset contains a total of 45,408 logs of appli-cations activities, it would require a large amount of time toidentify the usefulness of each application feature by employ-ing the whole dataset. As the purpose of this experiment wasmerely to identify relevant features and classifiers rather thandetermining the performance, a subdataset that was extractedfrom the telephony activities containing a total of 3,836 logsof calls from 20 randomly selected users was utilised. For theactual experiment, each of these 20 users data was dividedinto two halves based upon the date of initiation feature: thefirst half was used to generate a profile and the other half wasutilised to evaluate the performance of the classifiers and fea-tures in terms of false acceptance rate [FAR, false rejectionrate (FRR) and equal error rate (EER)] (e.g. the lower theEER, the better performance). The results of the preliminarystudy are presented in the following subsections.

    3.2.1 Radial basis function network

    An RBF neural network has been one of most popular pat-tern classification methods used in the Artificial Intelligence(AI) field [19]. By default, it has three network configura-tions: number of neurons, the performance goal and spread.For this preliminary study, only the number of neurons wasconfigured while the other two remained as default. Table 4presents the results achieved by the optimal RBF neural net-work configurations with several combinations of telephonyfeatures as the inputs. By employing the dialled telephonenumber and the location of call as the inputs with 75 neu-rons, the RBF network achieved an EER of 10.5 %.

    Table 4 The best RBF network configurations with various telephonyfeaturesFeatures No. of neurons FAR (%) FRR (%) EER (%)

    1, 2, 3, 4 75 13.9 15.8 14.91, 2, 4 100 13.7 15.3 14.51, 2, 3 50 13 13 131, 2 75 10.7 10.2 10.5

    Key: 1 telephone no; 2 location; 3 duration; 4 time

    Table 5 The best FF MLP network configurations with various tele-phony features

    Features No. of neurons FAR (%) FRR (%) EER (%)

    1, 2, 3, 4 125 11.9 30.6 21.31, 2, 4 100 11.5 29.7 20.61, 2, 3 150 10.5 40 24.71, 2 150 14.9 20.1 17.5

    Key: 1 telephone no; 2 location; 3 duration; 4 time

    3.2.2 Feed-forward multi-layered perceptron (FF MLP)network

    The FF MLP network is another widely employed AI tech-nique utilised in the pattern classification domain [19]. Withmore network configuration variables available, the FF MLPneural network is a more complex classifier compared withthe RBF neural network. For this study, apart from modify-ing the number of neurons, no other settings of the FF MLPneural network were adjusted. Table 5 presents the best exper-imental results obtained by several FF MLP network config-urations with various combinations of telephony features. Byusing the dialled telephone number and the location of thecall as the inputs and applying 150 neurons, the FF LMPclassifier achieved an EER of 17.5 %.

    3.2.3 A rule-based approach

    The basis for this approach was derived from the descriptivestatistics produced when analysing the data and the large vari-ances observed. A dynamic approach therefore seemed sen-sible to cope with the changing nature of the profile. Basedupon the premise that the historical profile can be utilisedto predict the probability of a current event, the rule-basedapproach illustrated in Eq. 1 was devised. The approach pro-vides a mechanism to ensure that all outputs are boundedbetween 0 and 1 to assist in defining an appropriate thresh-old. Alarm if:

    1 N


    Occurance of Featurei xMx=1 Occurance of Featurei x


    N threshold (1)


    i = The features of one chosen application (e.g. diallednumber for telephony application)

    x = The value of Featurei (e.g. office telephone numberand home telephone number)

    M = Total number of values for Featurei N = Total number of features Threshold = A predefined value according to each indi-

    vidual user


  • Active authentication for mobile devices utilising behaviour profiling

    Table 6 Experimental results by employing the rule-based approach

    Features FAR (%) FRR (%) EER (%)

    1, 2, 3, 4 18.3 21.9 20.11, 2, 4 8.1 16.7 12.41, 2, 3 17.8 21.7 19.71, 2 7.1 14.8 11

    Key: 1 telephone no; 2 location; 3 duration; 4 time

    For the experiment, all parameters within the formula werechosen according to the selected telephony features. Theresults from the experiment are presented in Table 6. By usingthe dialled telephone number and the location of the call, therule-based approach obtained an EER of 11 %.

    The above three sets of results have demonstrated that theusers can be discriminated by their telephony service withinthe mobile host environment with a good level of perfor-mance and also the usefulness of each application featuretowards the classification result. Both the dialled telephonenumber and location of call features proved to be positivefor discriminating the users by all three classifiers. In com-parison, neither the time nor the duration of call feature canbe considered as contributing positive information towardsthe classification process as by adding either or both of themas the inputs the performance of the classifiers decreased. Inaddition, the effectiveness of these features (i.e. location, tele-phone number, time and duration of call) was also suggestedby the previous statistical analysis presented in Sect. 3.1.

    Among the three chosen classifiers, the FF MLP neuralnetwork achieved the worst performance compared with theRBF neural network and the rule-based approach. Althoughthe RBF neural network had a slightly higher performance(0.5 % EER), it also consumed more than twice computa-tional power that the rule-based approach did (based uponobservations during the experimental study). As a result,the rule-based approach was employed as the classifier withwhich to progress the next step of the research as computingpower is a limited resource within the mobile environment.

    3.3 Behaviour profiling on mobile applications

    Based upon the findings from the descriptive statistics andpreliminary studies, a complete experiment was undertakenon mobile users application usage using the rule-basedapproach. In addition to the static profile technique, the exper-iment also utilised a dynamic profile technique to minimisethe impact caused by users irregular mobile usage, whichwas observed from the preliminary study (i.e. large FRRrates). The dynamic profile contains 7/10/14 days of eachusers most recent activities, updated on a daily basis. Theevaluation process for both static and dynamic profile tech-niques was carried out on the same dataset. A smoothing

    function that treats a number of successive applications activ-ities as one event was also introduced to cope with the mobileusers inconsistent and variable usage behaviour; therefore,a decision is made based upon the combined events ratherthan a single occurrence. The following sections contain acomplete behaviour profiling study on mobile users applica-tion usage considering standard, extended and multi-instanceapplications.

    3.3.1 Standard applications profiling

    For standard applications, the experiment employed all thestandard applications activities that are presented in Table 2.The application name, date of initiation and location of usagefeatures were extracted to generate profiles and validate theperformance of the classifier. A complete set of experimentalresults for users standard application usage is presented inTable 7. The best performance (EER 13.5 %) was obtainedby utilising a 14 day dynamic profile with the smoothingfunction of 6 application entries (also shown in Fig. 3).

    3.3.2 Extended applications

    The telephone call experiment employed all the telephonyactivities as described in Table 2. The telephone number, dateand location of call were extracted for the purpose of profilegeneration and data evaluation. A complete set of experimentresults for users telephone call usage is shown in Table 8.

    Table 7 Experimental results for standard applications

    1 (%) 2 (%) 3 (%) 4 (%) 5 (%) 6 (%)

    S 14 21.1 17.4 16.3 14.9 14.2 13.6D 14 21.1 17.3 16.0 14.5 14.0 13.5D 10 22.1 17.8 16.2 14.6 14.4 13.7D 7 24.0 19.4 17.6 15.9 15.3 14.4

    Key: S Static, D Dynamic

    Fig. 3 FAR-FRR plot for standard applications with the dynamic14 day profile with 6 application entries


  • F. Li et al.

    Table 8 Experimental results for the telephone call application

    1 (%) 2 (%) 3 (%) 4 (%) 5 (%) 6 (%)

    S 14 9.6 9.1 7.9 7.2 4.3 6.4D 14 8.8 8.1 6.4 6.4 6.3 5.4D 10 9.6 8.6 8.1 7.2 6.9 6.0D 7 10.4 8.8 8.5 7.3 7.0 6.2

    Key: S Static, D Dynamic

    Fig. 4 FAR-FRR plot for the telephone call application with thedynamic 14 day profile with 6 telephone call entries

    Table 9 Experimental results for the SMS application

    1 (%) 2 (%) 3 (%)

    S 14 9.6 9.1 7.9D 14 8.8 8.1 6.4D 10 9.6 8.6 8.1D 7 10.4 8.8 8.5

    Key: S Static, D Dynamic

    The best experimental result for the users telephony activityis an EER of 5.4 %, and it was achieved by using the 14 daydynamic profile technique with the smoothing function of 6telephone call entries (also depicted in Fig. 4).

    The SMS experiment utilised the SMS message data asdescribed in Table 2. For each SMS log entry, the texted tele-phone number, date and location of texting were extracted forbuild profiles and examine the performance of the classifier.As several participants had a limited number of messagesover the chosen 28 day period, a maximum of 3 log entrieswere utilised for the smoothing function. The complete resultfor users SMS application usage is shown in Table 9. The bestresult is an EER of 2.2 %, and it was acquired by the clas-sifier utilising the 14 day dynamic profile with a smoothingfunction of 3 text message entries (also shown in Fig. 5).

    In daily life, mobile users utilise their applications in achronological order. For instance, a user switches off thealarm clock (standard application) at 6:05 AM; at 7:10 AM,they make several phone calls (extended application) andstart listening to music (standard application) at 7:36 AM.

    Fig. 5 FAR-FRR plot for the SMS application with the dynamic 14 dayprofile with 3 text message entries

    Table 10 Experimental results for multi-instance applications

    1 (%) 2 (%) 3 (%) 4 (%) 5 (%) 6 (%)

    S 14 16.9 13.6 12.7 12.0 10.9 11.0D 14 16.5 13.5 12.1 11.6 10.5 10.1D 10 17.4 13.7 12.3 11.6 10.6 9.8D 7 19.0 15.2 13.1 12.4 11.3 10.5

    Key: S Static, D Dynamic

    Therefore, the multi-instance applications can continuouslypresent an image of what a user normally does, while eitherthe standard or extended applications could only partiallyprovide information on user activity. Hence, an experimentwas conducted to examine the performance of the multi-instance applications technique for constantly monitoringevery single activity to identify abnormal mobile usage.

    For the multi-instance applications experiment, all usersapplications activities that are presented in Table 2 wereutilised. For each user, their standard and extended appli-cations were joined together by using the time stamp in achronological order. Also, features were selected accord-ing to their application categories. In total, 30,428 stan-dard and 15,101 extended applications logs were employed.The experimental results for users multi-instance applica-tions activities are demonstrated in Table 10. By utilising thedynamic profiling technique with 10 days of profiling dataand a smoothing function with 6 log entries, the best EERresult of 10 % was obtained (also depicted in Fig. 6).

    3.3.3 Discussion

    In general, the dynamic profiling technique achieved aslightly better performance than the static profiling techniquedid. This is reasonable as a dynamic profile contains a usersmost recent activities rather than some usage that occurredseveral weeks ago and therefore is less relevant to userscurrent behaviour. Also, the performance is improved witha longer profile period. Hence, an increased number of days(e.g. 18/22 days) of user activities as a profile should be exam-


  • Active authentication for mobile devices utilising behaviour profiling

    Fig. 6 FAR-FRR plot for multi-instance applications with the dynamic10 day profile with 6 application entries

    ined to find the optimum solution, nonetheless [14], sug-gests that users only keep 67 % of new applications beyond a30-day period indicating that users do change their usagepattern overtime. While the smoothing function treated moreapplication entries as one incident, the performance was alsoimproved. The smoothing function reduces the impact anysingle event might have and seeks to take a more holisticapproach to monitoring for misuse; this will provide a user-friendly environment as fewer rejections occur when a userchanges their usage behaviour. However, this approach takesa longer time for the system to make a decision, providinga window of opportunity for misuse and a degree of abusemight be missed by the security control.

    The name of application and location of usage are valu-able features that can provide sufficient discriminatory infor-mation to individuate mobile users (as shown in Table 7).However, while this might identify many misuse scenarios,it would not necessarily identify all cases of misuse particu-larly those where a colleague might briefly misuse a devicebecause the location information is likely to fall within thesame profile as the authorised user. So, care is required ininterpreting these results. Limitations in the dataset are alsolikely to have created certain difficulties. Due to the smallnumber of mobile applications that were available in 2004,a large similarity of application usage between users withinthe dataset creates difficulty for any classification method.However, it is envisaged that current mobile users applica-tion usage would arguably differ more as there are more than1 million applications available in the mobile apps market.As a result, it would be arguably easier to discriminate mobileusers through their application usage now.

    The performance of telephony and the SMS applicationsis generally very good (as demonstrated in Tables 8 and 9)more than twice that of the standard application profiling.This reinforces the hypothesis that knowing both the appli-cation and what the user does with it improves the chance ofidentifying individual users significantly. Moreover, mobile

    Table 11 The performance comparison of behavioural techniques onmobile devicesBehavioural techniques EER (%)

    Behaviour profiling 10Gait recognition [11] 20.1Calls 4.2Keystroke analysis [5] 13Handwriting recognition [7] 1Voice verification [32] 7.8

    users had a far larger set of telephone contacts (the numbersthey can dial/text) than applications, making the classifica-tion process easier with more identifiable data points fromwhich to discriminate.

    The experimental results of the multi-instance application(as presented in Table 10) are between these from the standardand extended applications; this is within the expectation asthe experiment utilised the combination of both types. Basedupon all experimental results, it is envisaged that the largerthe proportion of extended applications users have, the betterthe performance of a system. As a result, the identificationprocess of which category (either standard or extended) anapplication belongs to is critical for a behaviour profilingsystem.

    The performance of the behaviour profiling technique iswithin the expectation of the overall behavioural biometriccategory in the mobile device environment (as illustrated inTable 11). However, in order to ensure this technique can bewidely adopted by users, it would have not only to provideadditional security but also user convenience. The next sec-tion of this paper describes a behaviour profiling frameworkthat is designed to continuously profile and verify mobileusers identity in a transparent style.

    4 A novel framework for behaviour profiling on mobiledevices

    The concept of transparently authenticating mobile deviceusers was first proposed by the Transparent Authentica-tion System (TAS), which utilises a mixture of biometrictechniques to verify a mobile users identity in a continu-ous and transparent manner [4]. Based upon the foundationlaid by TAS, the behaviour profiling framework provides anenhanced security for the mobile device with minimum userinconvenience and the framework works in the followingmanner:

    To improve the security for the mobile device beyond thatoffered by the password and token-based approach;

    To verify the user based upon their application usage in acontinuous style;


  • F. Li et al.

    To ensure the verification process is carried out in a user-friendly manner and that the user is mainly verified trans-parently rather than intrusively;

    To provide an architecture that can operate in one of threemodes based upon the desired output implementation: asa standalone security control, within an IDS system as amisuse detector or within a TAS.

    A number of process engines and a security manager havebeen devised to achieve these objectives (as illustrated inFig. 7). When a user utilises an application, details of theactivity are automatically collected by a data collectionengine and then formulated into a behavioural sample. Abehaviour classification engine compares the sample with aprofile that is pregenerated by a behaviour profile engine todetermine the legitimacy of the user. Based upon the verifica-tion result, the security manager can make one of the follow-ing decisions according to the mode in which the frameworkoperates: for the standalone mode, the security manager indi-vidually handles the result and responses accordingly; other-wise, the security manager forwards the result to a securitymanagement system (e.g. TAS) that makes any final deci-sions. A detailed description of this process is presented inthe following sections.

    4.1 Processing engines

    The main duty of the data collection engine is to capture amobile users application activities. When an application isutilised, the data collection engine automatically gathers fea-tures associated with that application based upon the infor-mation presented in the applications storage. The featurescan be either related to the system level information of theapplication (e.g. the name of the application and location of

    usage) or the personalised user data (e.g. telephone numbersand email recipients).

    The primary function of the behaviour profile engine isto generate various templates by utilising the combination ofthe users historical data, two template generation algorithmsfor standard and extended applications, a dynamic profilingtechnique and a smoothing function. The dynamic profilingtechnique is employed to maintain accuracy of the templates,while the smoothing function is utilised to provide a user-friendly environment.

    The behaviour classification engine provides the mainfunctionality for the verification process when a verifica-tion requirement is met: the smoothing function, a verifi-cation time or the sensitivity of an application. The criteriaof the smoothing function are that based upon the experi-mental results presented in Sect. 3, a verification will not beperformed unless a total of 6 applications have been utilised.The verification time is the time period that controls howlong the behaviour classification engine has to wait to per-form verification. Within the verification time, the require-ment of the smoothing function is utilised as the referencefor verification; otherwise, verification processes will be car-ried out regardless of the requirement of the smoothing func-tion; this forces the framework to always provide securitywithin the reference timeframe. The sensitivity of the appli-cation is a measure of the value of the application and/or itsdata. According to [22], a highly sensitive mobile applica-tion is associated with high-risk levels. When a user requestsaccess to a highly sensitive mobile application but the cur-rent security status of the device is below the requirementfor accessing it, the behaviour classification engine will ver-ify any applications which have not been verified and updatethe security status even though neither of the aforementionedrequirements are met.

    Fig. 7 A novel behaviour profiling framework


  • Active authentication for mobile devices utilising behaviour profiling

    Table 12 Application performance factor

    Application performance (EER) (%) Factor

    02 124 0.946 0.868 0.7810 0.61012 0.51214 0.41416 0.31618 0.218 and larger 0.1

    The communication engine acts as a gateway for theframework. By utilising it, the framework can downloadapplication features information from an online repositoryto facilitate the data collection process. When the frameworkoperates in standalone mode and the device is locked down,the communication engine sends a code to the user which theycan use to unlock their device. When the framework oper-ates in dependent mode, the communication engine works asa bridge between the framework and a comprehensive secu-rity management system by forwarding verification resultsand accepting commands to/from the security managementsystem.4.2 Security status module

    The main function of the security status module is to main-tain the system security status (SSS) that constantly indi-cates how secure the system is. By utilising the SSS level,the framework can provide or deny access to the user accord-ingly. The SSS level is a numeric value in the range of 3to +3:1 3 indicates low security while +3 indicates highsecurity; it is calculated based upon two critical factors: theperformance of an application and the verification result.The performance factor is dynamically allocated to eachapplication based upon their performance as demonstrated inTable 12.

    After the activity of an application is verified, a temporaryvalue will be allocated based upon its performance factor.When a verification process involves more than one applica-tion, the temporary value will be obtained by combining theperformance factor of each individual application. Depend-ing upon the verification result, the temporary value is thenadded to (verified successfully) or subtracted from (verifiedunsuccessfully) the existing SSS level to derive the currentSSS level as shown in Eq. 2.

    1 The boundaries defined on the numerical scale are only provided asa suggestion.

    SSS level =m

    i=1VR APFi + Existing SSS Level (2)


    VR = Verification Result APF = Application Performance Factor m = Number of applications

    The time that has elapsed between two verificationprocesses also affects the SSS level. It is important that theSSS level should be reduced when a device is not used fora period of time. In this way, the opportunity of a devicebeing abused would be reduced when the device is left witha high SSS level. The formula for degrading the SSS levelis illustrated in Eq. 3. The time period is an average dura-tion between applications being utilised by a mobile deviceuser. The time factor does not have any influence on a nega-tive SSS level. The negative SSS level can only be changedwhen a users behaviour is verified or an unlock code isentered.

    SSS level = Current SSS Level Time ElapsedTime Period


    where: Current SSS Level is >0

    4.3 Security manager

    The security manager is the central controller of the frame-work as it co-operates with other elements to complete var-ious tasks, including continuously verifying the users iden-tity, updating the performance factor of an application andmaintaining the SSS level. The key task of the security man-ager is to monitor the current SSS level and make subsequentdecisions when the user requests access to an application.This is achieved by utilising an SSS monitor and response(SMAR) algorithm, which is the core security component ofthe proposed framework (as illustrated in Fig. 8). The algo-rithm contains three main checking stages before the deviceis locked down. These checking stages were chosen to pro-vide a high level of user convenience and improved security.The algorithm employs a mixture of transparent and intrusivemethods to verify a users identity. However, it is envisagedthat the majority of legitimate users will experience transpar-ent phases, while intrusive verification challenges are onlyutilised to ensure a users legitimacy in the event of accessbeing required to the mobile device but the SSS level beingbelow security requirements.

    When a user requests access to an application, the securitymanager initially examines the sensitivity of the application(either highly sensitive or non-sensitive). A high-level secu-rity requirement is set for highly sensitive applications asthey are associated with critical information: they cannot beaccessed when the SSS level is below 2 unless the user passes


  • F. Li et al.

    Fig. 8 Security manager: SSS monitor and response algorithm

    an intrusive verification. In comparison, the security require-ment for non-sensitive applications is much more relaxed:they can be utilised when the SSS level is greater or equal to2; otherwise, an intrusive verification will be used beforethe application can be accessed.

    For the second checking stage, the security manager com-pares the current SSS level with the security requirements.If the requirements are met, the users will be granted accessto the application. Otherwise, further checking will be per-

    formed based upon the nature of the application. If theapplication is a non-sensitive application, the user will bechallenged with a randomly selected security question; ifthe application is associated with high sensitive informa-tion, the security manager will utilise the third verificationrequirement as described earlier and perform any subsequentactions.

    For the final checking stage, the security manager utilisesa randomly selected security question as the intrusive ver-


  • Active authentication for mobile devices utilising behaviour profiling

    ification method to verify the users identity. An intrusiveflag that indicates how many times the intrusive verificationmethod is utilised increases by 1 when the user answers thequestion; by default, the value of the intrusive flag is 0. Theuser will be granted access to the application if the questionis correctly answered. Also, the current SSS level will beincreased by a value (e.g. 1), and the intrusive flag will be setto 0. Otherwise, the intrusive flag remains the same, and theuser will be challenged again if the value of the flag is not>2. When the value of the flag equals 3, the device will belocked down and only an administrative security passwordcan unlock it. When a correct unlock code is entered, boththe value of the intrusive flag and the SSS level will be set to0 and the user will be able to access the device once again.

    The three-stage SMAR algorithm of the security man-ager is the core component of the framework to adjust thebalance between user convenience and system security. Fornon-sensitive applications, the user convenience is achievedby allowing users to have at least two transparent chances(depends upon the performance of individual applications) toverify themselves correctly before being intrusively authen-ticated. As a result, it is envisaged that if the system is work-ing correctly, the SSS level should be high enough to permitautomatic access for legitimate users, and their experienceof intrusive security challenges will be minimal. The sameconfiguration would also permit potential misuse on mobileapplications as long as the SSS level is above the threshold(i.e. not smaller than 2). This phenomenon is difficult toavoid for any behavioural-based biometric due to the trade-off between the FAR and the FRR. However, if the impostercontinuously abuses mobile services, the probability of mis-use that goes under the radar of the framework decreasesas eventually the SSS level will be reduced to the thresholdand the device will be locked down. For sensitive applica-tions, the chance of users being intrusively verified at thebeginning of usage is significantly higher compared withwhen they access non-sensitive applications. This intrusive-ness will be gradually reduced as legitimate users continuousactivities will keep the SSS level high enough to allow auto-matic access to sensitive applications. However, the chance ofimposters accessing sensitive applications is virtually impos-sible as their activities likely differ from a legitimate usersprofile forbidding the SSS level to satisfy the requirementsfor access to sensitive applications.

    When the framework works in the dependent mode,it can become a component for an authentication secu-rity mechanism (e.g. TAS) or an IDS security mechanism(e.g. the knowledge-based temporal abstraction (KBTA)method [27]) to complete these mechanisms and enhancetheir ability and performance. Therefore, the security man-ager only provides a verification result, and the final decisionwill be made by the other security mechanisms. As a result, itwould be difficult to evaluate the performance and impact of

    the framework on other security mechanisms when it oper-ates in the dependent mode.

    5 Evaluation

    In order to understand the effect that the framework has uponthe overall performance, two aspects of the framework shouldbe examined: the impact on the processing power and theeffectiveness of authenticating the user. Regarding consump-tion of the processing power, previous research demonstratesthat a complicated multi-modal biometric authentication sys-tem (i.e. TAS) was prototyped within the mobile environmentand users were satisfied with the performance [8]. There-fore, it is envisaged that the proposed framework will havea small processing power footprint and little effect on theperformance of the mobile device. For the performance ofauthentication, the framework was evaluated via a simula-tion process which was conducted within the Matlab environ-ment. The simulation system employed the same 76 users 4weeks mobile activities, which were utilised in Sect. 3 as thesimulation data. For each user, their activities were dividedinto two halves, containing first and second two-week activ-ities, respectively. A users profile was initially trained byutilising the first two-week worth of activities, with the pro-file then being updated on a daily basis. The rest of usersactivities was employed to evaluate the performance of thebehaviour profiling framework. Furthermore, due to the lackof sensitive application usage within this dataset, the textmessage application was selected as a sensitive applicationin order to evaluate the effect upon the framework. During thechosen period, 22 users utilised the text message application,representing 4.3 % of the total application usage.

    As discussed in Sect. 4, the performance of the frameworkcan be influenced by three key parameters: the smoothingfunction, the verification time and the degradation function.Therefore, the evaluation sought to analyse the effect theseparameters have upon the performance. As such, four sce-narios were set up to independently assess each parameter inturn, as illustrated in Table 13. Time periods for the degrada-tion function were set between 1 and 60 min with a 10-mininterval for all four scenarios.

    Table 13 Four scenarios for the simulation process

    Scenario Smoothing function Verificationtime (min)

    A 1 application NAB 3 applications 3C 3 applications 6D 6 applications 6


  • F. Li et al.

    5.1 Simulation results

    The framework was configured according to the setup inTable 13. Based upon the verification requirement of thebehaviour classification engine (in Sect. 4), a users appli-cation activity will be verified as soon as one application isutilised; therefore, the verification time is not applicable. Thesimulation results for scenario A are presented in Table 14.

    In comparison with the configuration of scenario A, sce-nario B employed more applications for the smoothing func-tion, allowing the smoothing function to work with up to 3application activities. According to the verification require-ments of the behaviour classification engine, within the 3 minverification time window, applications will be processedwhen the total number being utilised reaches 3. Otherwise,any utilised applications will be verified even when the totalnumber of them is smaller than 3. Table 15 demonstrates thesimulation results for scenario B.

    Based upon the setup of scenario B, scenario C utiliseda longer verification time; this increases the potential forallowing more application activities to be processed withinone smoothing function. Based upon the requirement of thebehaviour classification engine, application activities will beclassified as soon as the total number of them reaches 3 withinthe 6 min verification time window; when the 6 min verifica-tion time window is surpassed, even if the total number of

    Table 14 Simulation results of scenario ANon-sensitive Sensitive OverallFRR FAR FRR FAR FRR FAR

    1 min 9.05 4.35 98.89 0 12.91 4.1710 mins 8.97 4.35 98.89 0 12.53 4.1720 mins 8.88 4.36 98.89 0 12.19 4.1730 mins 8.82 4.36 98.89 0 12.01 4.1740 mins 8.71 4.36 98.89 0 11.77 4.1750 mins 8.66 4.36 98.89 0 11.59 4.1760 mins 8.6 4.36 98.89 0 11.45 4.17

    Table 15 Simulation results of scenario BNon-sensitive Sensitive OverallFRR FAR FRR FAR FRR FAR

    1 min 8.05 3.41 100 15.29 13.06 4.1710 mins 7.87 3.42 94.28 15.29 12.53 4.0820 mins 7.75 3.42 88.8 15.29 12.08 4.0830 mins 7.72 3.42 85.95 15.29 11.88 4.0840 mins 7.64 3.42 81.89 15.29 11.58 4.0850 mins 7.62 3.42 78.57 15.29 11.38 4.0960 mins 7.57 3.42 77 15.29 11.24 4.09

    application activities is smaller than 3, they will be processed.Simulation results of scenario C are presented in Table 16.

    In comparison with the setup for scenario C, scenario Demployed a higher number of applications for the smoothingfunction; this allows the smoothing function to potentiallywork with up to 6 application activities. According to the ver-ification requirements for the behaviour classification engine,within the 6 min verification time, application activities willbe classified once there are 6 applications being utilised.When the 6 min verification time is exceeded, applicationactivities will be processed by the behaviour classificationengine even though the total number of activities is

  • Active authentication for mobile devices utilising behaviour profiling

    88.55 % of the time the legitimate user will be transparentlyverified and automatically obtain access to the device. Withthe same configuration, the imposter has only got a 4.17 %chance to misuse an application and conversely 95.83 % ofthe time they will be denied access. It is worth noting that theabove simulation results did not include the intrusive stage ofthe authentication process. Therefore, in reality, it is highlylikely that the imposter will be intrusively prompted with arandomly selected security question, and hence, the devicewill be locked down, resulting in an improved overall FAR.Arguably, the behaviour profiling framework is capable ofproviding continuous and transparent protection for a goodproportion of the time and is able to do so in a more secureand user convenient fashion.

    The smoothing function, verification time and degrada-tion function are employed to justify the balance between theusers convenience and system security, and their impact wasalso examined through the simulation scenarios. As demon-strated by the simulation results (from Tables 14, 15, 16, 17),the best system performance (11.24 % for overall FRR and4.09 % for overall FAR) was obtained by utilising a com-bination of the smoothing function set to 3 applications, averification time of 3 min and degradation function of 60 min(scenario B). With other configurations, the overall systemperformance decreased slightly. What the simulation resultshave shown is the effect these parameters have upon theframework and the need to ensure that they are set on anindividual rather than population basis so that the system canbe configured to optimally perform given an individual usersbehaviour profile.

    6 Conclusions and future work

    The first part of this paper demonstrated a feasibility studywhich showed that by utilising a number of classifiers, mobiledevice users can be discriminated from each other based upontheir application usage. The rule-based approach proved themost suitable classifier for mobile device users in terms ofperformance and computational power. The dynamic profil-ing and smoothing techniques were also put in place to copewith the inconsistency of user behaviour, and as a result, bet-ter performance was obtained.

    Based upon the promising experimental result, the secondpart of this paper proposed the behaviour profiling frameworkand described its core components and functionalities. Theframework is able to provide a continuous and non-intrusiveverification mechanism in standalone, TAS or IDS modes.By monitoring users application activities, the SMAR algo-rithm can continuously evaluate the system security statusbased upon which the framework can make a decision ofwhether to grant access to users. The framework is sub-sequently analysed utilising a real user dataset in a set of

    scenarios, providing conclusive evidence that it can providetransparent identity verification for a good proportion of thetime, thereby outperforming traditional PIN-based authenti-cation. However, work is still required on improving the levelof accuracy.

    Future work will focus upon two directions. Firstly, upondeveloping a fully functioning prototype of the proposedframework so that a series of studies examining the prac-tical usability of the approach can be measured. Secondly,to undertake a wide-spread data collection activity so that amodern and relevant corpus of behavioural-based data can beproduced from which algorithms and models can be testedand evaluated.


    1. Bishop, M.: Neural Networks for Pattern Classification. OxfordUniversity Press, Oxford (1995)

    2. Boukerche, A., Nitare, M.S.M.A.: Behavior-based intrusion detec-tion in mobile phone systems. J. Parallel Distrib. Comput. 62(9),14761490 (2002)

    3. Buschkes, R., Kesdogan, D., Reichl, P.: How to increase securityin mobile networks by anomaly detection. In: Proceedings of the14th Annual Computer Security Applications Conference, pp. 312 (1998)

    4. Clarke, N.: Transparent User Authentication. Springer, Berlin(2011)

    5. Clarke, N.L., Furnell, S.M.: Authenticating mobile phone usersusing keystroke analysis. Int. J. Inf. Secur. 6(1), 114 (2006)

    6. Clarke, N.L., Furnell, S.M.: Authentication of users on mobiletelephonesa survey of attitudes and practices. Comput. Secur.24(7), 519527 (2005)

    7. Clarke, N.L., Mekala, A.R.: The application of signature recogni-tion to transparent handwriting verification for mobile devices. Inf.Manag. Comput. Secur. 15(3), 214225 (2007)

    8. Clarke, N.L., Karatzouni, S., Furnell, S.M.: Flexible and transpar-ent user authentication for mobile devices. In: Proceedings of the24th IFIP TC 11 International Information Security Conference,Pafos, Cyprus, May 1820, ISBN: 978-3-642-01243-3, pp. 112(2009)

    9. Credant, Phone Data makes 4.2 Million Brits Vulnerable to IDTheft, Credant, (2009), Accessed: 14 June 2012

    10. DARPA, Active Authentication, DARPA, (2011),Accessed: 17 April 2012

    11. Derawi, M.O., Nickel, C., Bours, P., Busch, C.: Unobtrusive userauthentication on mobile phones using biometric gait recognition.In: Proceedings of the 2010 Sixth International Conference onIntelligent Information Hiding and Multimedia Signal Processing,pp. 306311 (2010)

    12. Eagle, N., Pentland, A., Lazer, D.: Inferring social network struc-ture using mobile phone data. In: Proceedings of the National Acad-emy of Sciences (PNAS), vol. 106, pp. 1527415278 (2009)

    13. FBI, Smishing and Vishing, The FBI, (2010),Accessed: 11 April 2012

    14. Flurry, Mobile Apps: Models, Money and Loyalty. FlurrySmartphone Industry Pulse,


  • F. Li et al.

    Mobile-Apps-Models-Money-and-Loyalty (2009), Accessed: 01August 2012

    15. Gosset, P.: ASPeCT: Fraud Detection Concepts: Final Report. DocRef. AC095/VOD/W22/DS/P/18/1 (1998)

    16. Hall, J., Barbeau, M., Kranakis, E.: Anomaly based intrusion detec-tion using mobility profiles of public transportation users. In: Pro-ceeding of IEEE International Conference on Wireless and MobileComputing, Networking and Communications, vol. 2, pp. 1724(2005)

    17. ITU, Key Global Telecom Indicators for the World Telecommu-nication Service Sector, International Telecommunication Union,, Accessed 01 April 2012

    18. Jain, A.K., Duin, R.P.W., Mao, J.: Statistical pattern recognition: areview. Pattern Anal. Mach. Intell. IEEE Trans. 22(1), 437 (2000).doi:10.1109/34.824819

    19. Jain, A.K., Mao, J., Mohiuddin, K.M.: Artificial neural networks:a tutorial. Computer 29(3), 3144 (1996). doi:10.1109/2.485891

    20. Kaspersky Lab, European Users Mobile Behaviour and Awarenessof Mobile Threats, Kaspersky Lab ZAO, (2011), Accessed: 25 May 2012

    21. Kurkovsky, S., Syta, E.: Digital natives and mobile phones: A sur-vey of practices and attitudes about privacy and security. In: Pro-ceedings of the IEEE International Symposium on Technology andSociety (ISTAS), pp. 441449 (2010)

    22. Ledermuller, T., Clarke, N.L.: Risk assessment for mobile devices.In: Proceedings of Privacy and Security in Digital Business 8thInternational Conference, TrustBus, pp. 210221 (2011)

    23. Power, R.: Mobility and Security: Dazzling Opportunities, Pro-found Challenges, McAfee, (2011), Accessed: 1 May2012

    24. Samfat, D., Molva, R.: IDAMN: an intrusion detection architecturefor mobile networks. IEEE J. Sel. Areas Commun. 15(7), 13731380 (1997)

    25. Samsung., Galaxy Nexus, Samsung, (2012), Accessed: 04 May 2012

    26. Securelist, Mobile Malware Evolution: An Overview,Part 3, Securelist, (2009), Accessed: 30 March 2012

    27. Shabtai, A., Kanonov, U., Elovici, Y.: Intrusion detection for mobiledevices using the knowledge-based, temporal abstraction method.J. Syst. Softw. 83(8), 15241537 (2010)

    28. Sun, B., Chen, Z., Wang, R., Yu, F., Leung, V.C.M.: Towards adap-tive anomaly detection in cellular mobile networks. In: the IEEEConsumer Communications and Networking Conference, vol. 2,pp. 666670 (2006)

    29. Sun, B., Yu, F., Wu, K., Leung, V.C.M.: Mobility-based anomalydetection in cellular mobile networks. In: Proceedings of ACMWireless Security (WiSe 04), pp. 6169 (2004)

    30. Which?, 13.5 million UK mobile phone users at risk offraud, Which? Tech Daily, (2011), Accessed: 31 July 2012

    31. Wolpert, D.H., Macready, W.G.: No free lunch theorems for opti-mization. IEEE Trans. Evolut. Comput. 1, 6782 (1997)

    32. Woo, R., Park, A., Hazen. T.: The MIT mobile device speakerverification corpus: data collection and preliminary experiments.In: Speaker and Language Recognition Workshop, pp. 16 (2006)


    Active authentication for mobile devices utilising behaviour profilingAbstract 1 Introduction2 Behavioural biometrics for mobile devices3 Behavioural profiling on mobile devices3.1 Descriptive statistics analysis3.2 A preliminary study utilising telephony activity3.2.1 Radial basis function network3.2.2 Feed-forward multi-layered perceptron (FF MLP) network3.2.3 A rule-based approach

    3.3 Behaviour profiling on mobile applications3.3.1 Standard applications profiling3.3.2 Extended applications3.3.3 Discussion

    4 A novel framework for behaviour profiling on mobile devices4.1 Processing engines4.2 Security status module4.3 Security manager

    5 Evaluation5.1 Simulation results5.2 Discussion

    6 Conclusions and future workReferences


View more >