Analisis penekanan kunci dinamik untuk verifikasi biometrik berbasis jaringan syaraf tiruan propagasi balik

Abstract

ARIF BUDIMAN. Dynamic keystroke analysis for biometric verification based on artificial neural network backpropagation. Under the direction of SUGI GURITMAN and HERU T. NATALISA. Biometric keystroke analysis studies the typing pattern of human behavior. Literature journals show a good prospect for a verification system and commonly use static method (where training data is similar with the verification data). This research uses dynamic analysis method. The data will be analyzed into certain dominant character pairs so that it is possible for training data to be different than the verification data (using random data display). The dominant character pairs were obtained from preliminary research which was conducted to specifically analyze 756 words in the Indonesian language. It is based on left-right finger movement. The character pairs are an, ng, la, en, ka while the cluster of close character pairs are (ng,ba,ma,na,nd), (an,am,ab), (su,au,ay,gu,di,du,di,ai), (pa,is,ia,ya,ua,us), and (ri,ti,ru,tu,ro,to,ep). The system prototype uses three variables, namely duration (d), interkey (i) and total time (T). Each variable has its own separated Artificial Neural network (ANN) by pre-processing input into five fuzzy classes (SC, C, S, L, SL). The training algorithm is backpropagation and uses two types of training models: A (crispy) and B (ambiguity). The training data were obtained by statistical analysis (mean, median, modus, standard deviation, minimum, maximum) and from fuzzy class pattern complementary data for each character pairs. The research was conducted to study the effect of keyboard layout on False Rejection Rate (FRR) percentage, the effect of data re-taken on % FRR, the effect of ANN verification test on the responses to the new exclusive data (which had never been included in training process). It also studies the effects of identification tests in terms of False Acceptance Rate (FAR) percentage on ANN training model A and B, and the effect of overall response in % FRR on intrusion detection. The conclusions are that model A gives lower % FRR and higher % FAR than model B while model B has a more flexible value of the decision threshold to be used as fuzzy output model in future research. The system stability is affected significantly by keyboard layout and environment conditions during typing. However, the data re-take process carried out more than 56 days afterwards still gave consistent results. ANN has the capability to recognize exclusive data with relatively lower % FRR results. Intrusion can be detected easily by spotting the existence of the % FRR spike. This model has a good capability for Biometric keystroke verification with dynamic text typing but still requires more research for Biometric identification.