104
Internet of Things (IoT)
The methods were used on keystroke dynamics datasets, where statistical measurements are common like mean, median, and standard deviation. Many distance-based
algorithms were also used as pattern recognition techniques such as Canberra, Chebyshev,
Czekanowski, Gower, Intersection, Kulczynski, Lorentzian, Minkowski, Motyka, Ruzicka,
Soergel, Sorensen, Wavehedges, Manhattan Distance, Euclidean Distance, Mahanobolis
Distance, Z Score, and KMean. Some machine learning methods also have been applied like
support vector machine (SVM), naïve bayes, multi layer perceptron, fuzzy set, K-nearest
neighbor, OneR, hidden Markov model (HMM), Gaussian Markov model (GMM), and random forest. Other methods like direction similarity measure (DSM), degree of disorder,
and array disorder were also used. ACO, PSO, Best First, and GA were used as optimization techniques.
Many classification methods have been applied over the last 30 years. Where statistical methods were common, over the last 10 years, strong machine learning and distancebased methods are common approaches. Figure 6.4 indicates the percentage distribution
of the type of methods used in literature.
To test the biometric system, few parameters (EER, FAR, FRR, etc.) are used to evaluate
the performance of the system. Figure 6.5 indicates the percentage distribution of average
EER previously recorded. The European standard for access control specifies that FAR
must be less than 1% and FRR must not be more than 0.001% [16]. But in literature, only
1.36% of studies [17,18] reached those acceptable results. In [17], large samples were collected from each subject in training session which is impractical in real life, whereas Ali
and Salami [18] used only key pressure as keystroke features and took data from only
seven subjects, with the scalability of the study being very low. Hence, further research has
to be done on keystroke dynamics for identification/verification of users since the application area of this technique is very large.
To summarize, most of the works have been studied on datasets collected through a
computer keyboard than a touch screen device. Many researchers have created the dataset
and applied classification algorithms and obtained the results. Some of the researchers
worked on optimization techniques and endeavored to enhance the keystroke dynamics
user authentication performance. Giot et al. [7] extracted the gender feature only from the
typing pattern on computer keyboard, but did not work on the typing pattern on touch
screen. They also used this soft biometric feature with the timing features and obtained
gain accuracy of up to 20%.
Distance
based
22%
Others
11%
Statistical
14%
Machine
learning
53%
FIGURE 6.4
Percentage distribution of different classification methods used in literature.
Internet of Things (IoT)
The methods were used on keystroke dynamics datasets, where statistical measurements are common like mean, median, and standard deviation. Many distance-based
algorithms were also used as pattern recognition techniques such as Canberra, Chebyshev,
Czekanowski, Gower, Intersection, Kulczynski, Lorentzian, Minkowski, Motyka, Ruzicka,
Soergel, Sorensen, Wavehedges, Manhattan Distance, Euclidean Distance, Mahanobolis
Distance, Z Score, and KMean. Some machine learning methods also have been applied like
support vector machine (SVM), naïve bayes, multi layer perceptron, fuzzy set, K-nearest
neighbor, OneR, hidden Markov model (HMM), Gaussian Markov model (GMM), and random forest. Other methods like direction similarity measure (DSM), degree of disorder,
and array disorder were also used. ACO, PSO, Best First, and GA were used as optimization techniques.
Many classification methods have been applied over the last 30 years. Where statistical methods were common, over the last 10 years, strong machine learning and distancebased methods are common approaches. Figure 6.4 indicates the percentage distribution
of the type of methods used in literature.
To test the biometric system, few parameters (EER, FAR, FRR, etc.) are used to evaluate
the performance of the system. Figure 6.5 indicates the percentage distribution of average
EER previously recorded. The European standard for access control specifies that FAR
must be less than 1% and FRR must not be more than 0.001% [16]. But in literature, only
1.36% of studies [17,18] reached those acceptable results. In [17], large samples were collected from each subject in training session which is impractical in real life, whereas Ali
and Salami [18] used only key pressure as keystroke features and took data from only
seven subjects, with the scalability of the study being very low. Hence, further research has
to be done on keystroke dynamics for identification/verification of users since the application area of this technique is very large.
To summarize, most of the works have been studied on datasets collected through a
computer keyboard than a touch screen device. Many researchers have created the dataset
and applied classification algorithms and obtained the results. Some of the researchers
worked on optimization techniques and endeavored to enhance the keystroke dynamics
user authentication performance. Giot et al. [7] extracted the gender feature only from the
typing pattern on computer keyboard, but did not work on the typing pattern on touch
screen. They also used this soft biometric feature with the timing features and obtained
gain accuracy of up to 20%.
Distance
based
22%
Others
11%
Statistical
14%
Machine
learning
53%
FIGURE 6.4
Percentage distribution of different classification methods used in literature.