4 Propuesta
4.2 BASE FILOSÓFICA
Summary of the state of the art research:
• Most of the research is conducted in a controlled environment or with a predefined task. A controlled environment or predefined task does not represent the real world scenario for CA, due to the fact that users are concentrated towards completion of the tasks, which might influ- ence their normal behaviour.
• Except [19], all of these researches were conducted in either periodic manner or in a traditional authentication manner with continuous data. In our work we will focus on an actual CA
2. STATE OF THEART
system where each and every performed action by the user will be taken into consideration and according to the genuineness of the performed action(s) the system will decide whether the present user may continue to work or not.
• Most of these works represent their system performance in terms of EER of FMR and FNMR. According to our understanding this is not the proper way to report the CA system perfor- mance. We will provide an alternative CA system performance measure metric in our research (see Chapter 4).
Part II
Chapter 3
Trust Model: A Computational Approach for
Continuous Autentication
In this chapter, we describe the approach followed for CA. As we are trying to build a CA sys- tem that reacts immediately on every action or activity performed by the user. Therefore, a robust computational algorithm is required, that can handle the large variation of the user’s behaviour. We developed a trust function that will be described in this chapter.
This chapter is based on the paper published in: [95] MONDAL, S.,ANDBOURS, P. A com- putational approach to the continuous authentication biometric system. Information Sciences 304 (2015), 28 – 53.
3.1
Introduction
The basic idea is that the trust or confidence of the system in the genuineness of the current user depends on the deviations from the way this user performs various actions on the system. If a specific action is performed in accordance with how the genuine user would perform the task (i.e. as it is stored in the genuine user’s profile/template), then the system’s trust in the genuineness of this user will increase, which is called Reward. If there is a large deviation between the behaviour of the genuine user and the current user, then the trust of the system in that user will decrease, which is called Penalty. The amount of change of the trust level can be fixed or variable. A small deviation from the behaviour of the user, when compared to the template, could lead to a small decrease in trust, while a large deviation could lead to a larger decrease.
No single person will be able to always behave in exactly the same manner [15]. For the genuine user this means that he/she will also sometimes deviate more from his/her normal behaviour, which will lead to a decrease in trust. However, the majority of actions that a genuine user will perform will be close to his/her normal behaviour, i.e. lead to an increase of trust. Overall this would lead to a high level of trust. For an imposter however the opposite holds. In some cases he will behave as the genuine user, increasing his level of trust, but the majority of actions will lead to a decrease in trust due to the large deviation from the behaviour of the genuine user. This will then lead to a general decrease of the trust over time for an imposter user. Obviously an ideal system should perform in such a way that the trust in anyone other then the genuine user will decrease fast to a value below a predefined threshold Tlockout, then the system locks itself and will require static authentication
of the user to continue working. In such an ideal system, also a genuine user would never reach a trust level that would result in a lockout, i.e. the genuine user would not notice the presence of the continuous authentication system in his daily activities.
In Figures 3.1 and 3.2 we elaborate the concept of Trust Model in more details. In Figure 3.1 we see how the trust level changes when we compare the profile of a genuine user with test data of that genuine user. In this figure, we see that the trust level, sometimes decreases due to penalties, but it never drops below the lockout threshold (i.e. Tlockout= 90, marked by a red line). Figure 3.2 shows
that if the same user’s profile is compared to test data of an imposter user, the trust will drop (in this example) 5 times below the lockout threshold within 500 user actions1. We set the upper limit of the
system trust is 100, to prevent a situation where an imposter user benefits from the high system trust obtained by the genuine user, before he/she hijacks the system.
3. TRUSTMODEL: A COMPUTATIONALAPPROACH FORCONTINUOUSAUTENTICATION 0 50 100 150 200 250 300 350 400 450 500 85 90 95 100 Event Number System Trust T lockout
Figure 3.1: Trust value for genuine user tested with the genuine test data.
0 50 100 150 200 250 300 350 400 450 500 85 90 95 100 N1 N2 N3 N4 N5 Event Number System Trust T lockout
Figure 3.2: Trust value for genuine user tested with the imposter test data.
Bours [19] has described a trust model concept for continuous authentication using keystroke dynamics. He demonstrated that the trust level will increase or decrease according to the absolute distance (i.e. scaled Manhattan distance) between the genuine user’s template and the current typing. We can also use classifier scores (i.e. the probability of the genuineness of that event or action) to increase or decrease the trust value. There are two major concerns we need to address related to the trust model:
• The threshold for the Penalty or Reward: The classifier score (i.e. sci= P (x|H1) where xiis
the feature vector of the ithperformed action and H1is the hypothesis for the genuine user)
generally ranges from 0 to 1. Therefore, we have to set a threshold (T r) where, sci≥ T r we
will give Reward otherwise Penalty.
• The amount of the Penalty or Reward the system will give corresponding to the classifier score:The amount of penalty or reward can be fixed, for example if sci≥ T r increase the
3.2 STATICTRUSTMODEL
Algorithm 3.1: Algorithm for 3-level Static Trust Model. Data:
sci← Classification score for the ithaction
T r ← Threshold for Penalty or Reward T rP← Threshold for 2ndlevel Penalty
T rusti−1← System trust after (i − 1)thaction
Result:
T rusti→ System trust on the user after ithaction 1 begin 2 if sci≥ T r then 3 ∆T(sci) = fReward(sci) 4 else 5 if T rP ≤ sci< T r then 6 ∆T(sci) = −fP enalty1 (sci) 7 else 8 ∆T(sci) = −fP enalty2 (sci)
9 T rusti= min {max {T rusti−1+ ∆T(sci) , 0} , 100}
trust value by 1 otherwise decrease the trust value by 1. On the other hand, the amount of penalty or reward can be variable depending upon the actual scivalue.
Based on the above mentioned criteria different computational algorithms can be implemented in the Trust Model. Some of these are explained below.