Interactive Role Mining Including Expert Knowledge into Evolutionary

Algorithms

Simon Anderer

, Nicolas Justen

, Bernd Scheuermann

and Sanaz Mostaghim

SIVIS GmbH, Gr

unhutstrasse 6, 76187 Karlsruhe, Germany

Karlsruhe University of Applied Sciences, Karlsruhe, Germany

Otto-von-Guericke Universit

at Magdeburg, Magdeburg, Germany

Keywords:

Dynamic Role Mining, User Interaction, Evolutionary Algorithms.

Abstract:

To protect the security of corporate IT systems against erroneous or fraudulent behavior of employees, Role

Based Access Control has proven to be an effective concept. The corresponding NP-complete Role Mining

Problem aims at ﬁnding a minimal set of roles and an assignment of roles to users. A valuable source of

additional information, which is not yet included in current role mining algorithms, is expert knowledge. Users

of role mining software should be enabled to monitor the role mining process and interactively transfer their

knowledge to the system, for example by suggesting good or deleting bad roles. This leads to dynamically

occurring interaction events, which must be included into the optimization process preferably in real-time,

since these have the potential to accelerate the optimization process or improve the solution quality. This

paper introduces to interactive role mining. For this purpose, the hitherto static RMP is considered as dynamic

optimization problem. Since evolutionary algorithms have shown to be a promising solution approach, it is

shown how events emerging from user interaction can be integrated. The integration of different interaction

events into the evolutionary algorithm and their impact on the optimization process are then evaluated in a

range of experiments.

1 INTRODUCTION

IT systems of companies and organizations, in which

several people collaborate, must be protected not only

against external attacks such as malware and phish-

ing, but also against erroneous or fraudulent employee

behavior. According to a study by Pricewaterhouse-

Coopers, more than half of all fraud cases, that could

be identiﬁed within the participating companies, can

be traced back either exclusively to internal perpe-

trators or to a combination of internal and external

perpetrators (PwC, 2022). One approach to secure

IT systems against such threats is Role Based Ac-

cess Control (RBAC). At this, permissions, which

correspond to the authorizations necessary to perform

an operation on a data or business object, are not

assigned to the users of an IT system directly, but

are grouped into roles, which are then assigned to

users (Ferraiolo and Kuhn, 1992). This process is

also called role engineering and is carried out either

in a top-down or bottom-up fashion. The top-down

approach, in which company structures and business

processes are analyzed in order to derive suitable

roles, is very time-consuming and requires substan-

tial human effort. Therefore, the bottom-up approach,

in which role concepts are generated automatically,

has gained more and more attention in recent years.

The corresponding optimization problem is called the

Role Mining Problem (RMP) and was shown to be

NP-complete (Vaidya et al., 2007). It aims at ﬁnd-

ing a minimum number of roles based on a given as-

signment of permissions to users in order to enhance

comprehensibility and manageability. One source of

information, which can be of great value for the au-

tomatic creation of role concepts, but has not been

considered in previous role mining literature, is ex-

pert knowledge. In practice, RBAC implementation

projects are usually accompanied by experts to en-

sure that the implemented role concept meets estab-

lished security standards and compliance rules. These

experts have the ability to determine the suitability

of certain roles included in role concepts depending

on the company structure or the given assignments

of permissions to users. Hence, experts should be

able to monitor the role mining process, analyze the

proposed role concepts and interactively suggest im-

Anderer, S., Justen, N., Scheuermann, B. and Mostaghim, S.

Interactive Role Mining Including Expert Knowledge into Evolutionary Algorithms.

DOI: 10.5220/0012153000003595

In Proceedings of the 15th International Joint Conference on Computational Intelligence (IJCCI 2023), pages 151-162

ISBN: 978-989-758-674-3; ISSN: 2184-3236

151

provements. This results in dynamically occurring in-

teraction events, which should be integrated into the

optimization process, if possible, in real-time.

This paper aims at providing a framework to in-

clude expert knowledge into evolutionary algorithms

(EAs). First, the Role Mining Problem is introduced

in Section 2. In Section 3, the application of EAs

to the RMP is discussed. Special attention is given

to the addRole-EA, an evolutionary algorithm for the

Basic RMP, which, due to its role-centric approach,

has been selected as basis for the evaluation of differ-

ent interaction events. A detailed overview of user in-

teraction options in the context of role mining is pro-

vided in Section 4. Section 5 describes how dynam-

ically occurring interaction events can be integrated

into EAs in close to real-time. In Section 6, the in-

clusion of interaction events and their potential to en-

hance the optimization process is examined in differ-

ent evaluation scenarios. Section 7 concludes the pa-

per and presents paths for future research.

2 THE ROLE MINING PROBLEM

In the following, the main elements of the RMP are

introduced:

• U = {u

, u

, ...,u

} a set of M = |U| users.

• P = {p

, p

, ..., p

} a set of N = |P| permissions.

• R = {r

, r

, ..., r

} a set of K = |R| roles.

• UPA ∈ {0, 1}

M×N

the targeted permission-to-user

assignment matrix, where UPA

i, j

= 1 implies that

permission p

is assigned to user u

• UA ∈ {0, 1}

M×K

a role-to-user assignment matrix.

• PA ∈ {0, 1}

K×N

a permission-to-role assignment

matrix.

• RUPA := UA ⊗ PA ∈ {0, 1}

M×N

the result-

ing permission-to-user assignment matrix, where

⊗ denotes the Boolean Matrix Multiplication:

(UA ⊗ PA)

i, j

= max

l∈{1,...,k}

(UA

i,l

· PA

l, j

• π := hR,UA, PAi a role concept.

• Π the set of all role concepts.

Based on this, the RMP can be deﬁned as a matrix

decomposition problem:

Basic Role Mining Problem. Given a set of users

U, a set of permissions P and a permission-to-user

assignment matrix UPA, ﬁnd a role concept π com-

prising a minimal set of Roles R, a corresponding

role-to-user assignment matrix UA and a permission-

to-role assignment matrix PA, such that each user is

assigned exactly the permissions granted by UPA.

Basic RMP =

(

min |R|,

s.t. d(UPA, RUPA) = 0,

(1)

where d(A, B) := kA − Bk denotes the distance

of two matrices A, B ∈ R

m×n

and kAk denotes

the sum of absolute values of elements of A:

kAk :=

∑

i=1

∑

j=1

i, j

A role concept π is called a feasible solution for

the Basic RMP, if it satisﬁes the constraint in (1). In

this case, π is also denoted 0-consistent.

Since the RMP is a well-known optimization

problem, a variety of established solution strategies

has been proposed. Many approaches are based on

the concept of permission grouping. Permissions are

grouped into a set of candidate roles. These candi-

date roles are then assigned to users, mostly using

greedy approaches (Blundo and Cimato, 2010; Mol-

loy et al., 2009; Vaidya et al., 2010b). Another ap-

proach consists of mapping the RMP to other known

problems in data science, e.g. the Minimum Tiling

Problem (Vaidya et al., 2010a), the Minimum Bi-

clique Cover Problem (Ene et al., ) or the Set Cover

Problem (Huang et al., 2015), and using correspond-

ing olution approaches. A detailed overview of dif-

ferent solution strategies is provided by Mitra (Mitra

et al., 2016).

3 EVOLUTIONARY

ALGORITHMS FOR THE RMP

In recent years, the application of evolutionary algo-

rithms (EAs) in the context of the RMP has gained

more and more attention. Since these are particu-

larly well-suited for the integration of events emerg-

ing from user interaction, their application in the con-

text of the RMP is considered in more detail at this

point.

Each individual of an EA for role mining repre-

sents a possible solution of the RMP and thus a role

concept. Therefore, its chromosome consists of a UA

and a PA matrix. The set of roles R can be obtained

from the rows of PA. An important differentiation cri-

terion for EAs in the context of role mining is com-

pliance with the 0-consistency constraint. Most role

mining approaches based on EAs employ standard

crossover and mutation methods such as one-point

crossover and bitﬂip mutation. These lead to ran-

dom changes in UA and PA, which correspond to ran-

dom assignments of permissions to roles or of roles

ECTA 2023 - 15th International Conference on Evolutionary Computation Theory and Applications

152

to users, such that the 0-consistency constraint can

be easily violated. Therefore, the corresponding EAs

are not applicable for the Basic RMP (Saenko and

Kotenko, 2011; Saenko and Kotenko, 2012; Saenko

and Kotenko, 2016b; Du and Chang, 2014).

In (Anderer et al., 2020), the addRole-EA is pre-

sented, which comprises new, role-centric operators

for crossover and mutation, that ensure compliance

of the individuals of the addRole-EA with the 0-

consistency constraint at all times within the opti-

mization process. The role-centric approach of the

addRole-EA is particularly well-suited for the inte-

gration of the interaction events considered in Sec-

tion 6. A top level description of the addRole-EA

including the values of its parameters is provided in

Figure 1. Thereafter, the different components and

methods of the addRole-EA are described.

Post-

Processing

Selec on

Crossover

+ Muta on

Replacement

Pre-

Processing

Ini aliza on

+ Evalua on

START

END

Stop?

Evalua on

Popula on Size: 20 Crossover Rate: 0.1 Muta on Rate: 1.0

Eli sm Rate: 0.7 (SC1): 100,000 (SC2): 10,000

Yes

Figure 1: Integration of interaction events into addRole-EA.

Pre- and Post-Processing. In a ﬁrst step of the

addRole-EA, the size of the UPA matrix represent-

ing the RMP is reduced. For this purpose, four pre-

processing steps have been identiﬁed comprising for

example the removal of users that are assigned the

same set of permissions except for one representative.

The post-processing step adapts the obtained role con-

cepts to the original problem size.

Initialization. In order to create an initial popula-

tion, ﬁrst a seed individual is created from UA = U PA

and PA = I

, where I

denotes the N-dimensional

identity matrix. Since UA ⊗ PA = UPA ⊗ I

= UPA,

it complies with the 0-consistency constraint. Subse-

quently, according to the desired population size, new

individuals are created from the seed individual by ap-

plying a random series of mutation operators.

Evaluation. Since the addRole-EA is an algorithm

speciﬁcally tailored to the Basic RMP, the ﬁtness of

an individual equals the number of roles of the repre-

sented role concept.

Selection, Crossover and Mutation. At ﬁrst, the

individuals for crossover and mutation are selected

randomly. Within crossover, roles of two individuals

are exchanged to create offspring. Within mutation,

new roles are created and added to the chromosome

of an individual. In both cases the addRole-method,

which constitutes the main method of the addRole-

EA, is used to add a new role to the chromosome of

an individual. At this, not only a new role is added,

but the role structure is analyzed to identify and delete

roles that have become obsolete through the addition

of the new role. Figure 2 provides an example of the

addRole-method. For better visibility, ones are dis-

played as black dots in UPA, UA and PA. Zeroes are

represented as white dots.

UPA

+ new Role

Individual

Star ng Point

UA PA

Addi on of new role Removal of obsolete roles

UA PA

1 2

Figure 2: Example of addRole-method.

In this example, a new role, which is assigned per-

missions p

, p

and p

, is added to the individual.

For this purpose, a new row is appended to PA rep-

resenting the new role. In the ﬁrst step, the new role

is assigned to all users, that are assigned at least the

same permissions. These are users u

and u

. A new

column is appended to UA representing the assign-

ments of the new role to users. In the second step, it

is checked, whether a role has become obsolete. For

role r

, which assigns p

and p

to u

and u

, these

assignments are now also covered by the new role.

The same is valid for the assignments of permissions

to users induced by r

. Hence, these roles are deleted

from the individual such that in total the number of

roles could be reduced by one.

Replacement. The addRole-EA is a steady-state

evolutionary algorithm. For replacement an elitist se-

lection scheme is applied. First, based on the elitism

rate, individuals for the next generation are selected

based on their ﬁtness. The remaining individuals are

then selected randomly to ensure diversity among the

individuals of the next generation’s population.

Stopping Condition. The addRole-EA is termi-

nated according to two criteria: either a maximum

number of iterations has been executed (SC1), or no

further improvement has been achieved within a given

number of iterations (SC2).

Interactive Role Mining Including Expert Knowledge into Evolutionary Algorithms

153

4 USER INTERACTION IN ROLE

MINING

Although it seems natural to integrate expert knowl-

edge into ongoing bottom-up role mining processes,

to enhance the quality of the resulting role concepts,

or to accelerate the optimization process, to the best

of our knowledge no research has been conducted in

this area so far. Some approaches consider the inte-

gration of dynamically occurring events into the role

mining process. However, these events stem from

changes in the company structure, such as new em-

ployees joining a company and employees leaving a

company (Anderer et al., 2021a), or from changes in

the assignment of permissions to users (Saenko and

Kotenko, 2016a). Events caused by user interaction

are not considered.

onig and Schneider examine possibilities of how

users of optimization software can interact with an

optimization algorithm used for automatic generation

of layouts. In order to avoid confusion with regu-

lar users of IT systems in the context of role mining,

such users of optimization software will be referred

to as decision makers (DM) throughout the remainder

of this paper. Considering automatic layout genera-

tion, DMs are provided with the option to move, scale,

add and remove elements during the layout genera-

tion process. Furthermore, they distinguish between

direct and indirect manipulations. Direct manipula-

tions correspond to a modiﬁcation of solution candi-

dates. Indirect manipulations include the modiﬁca-

tion of optimization objectives or constraints as well

as the adaptation of parameters of the applied opti-

mization algorithm (K

onig and Schneider, 2011).

Nascimiento examined the interaction of a DM

with evolutionary algorithms for different graph-

based problems. For the Graph Clustering Problem,

the DM is given the option to either destroy clusters

or merge two clusters. For the Graph Drawing Prob-

lem, the DM is provided with the possibility to move

vertices. For the Map Labeling Problem, the DM is

given the option to exchange two vertices. Further-

more, the DM is allowed for dynamic focusing of the

optimization on manually chosen sub-problems. Fi-

nally, interaction possibilities aiming at the speciﬁ-

cations of evolutionary algorithms, like deliberate in-

clusion of certain individuals into the population of

an evolutionary algorithm are presented (Do Nasci-

mento, 2003).

Most of the current research focuses on the area

of indirect interactions. The goal is to help the EA

ﬁnd the regions of interest on the pareto front for the

DM. For this purpose many different methods have

been developed in the past. An overview is provided

Table 1: Interaction events in the context of role mining.

ID Event

I01 Add a new role

I02 Delete an existing role

I03 Merge roles (Create a new role from two existing roles)

I04 Split roles (Create two new roles from one existing role)

I05 Edit PA (whilst preserving 0-consistency)

I06 Edit UA (whilst preserving 0-consistency)

I07 Adapt mutation rate

I08 Adapt crossover rate

I09 Adapt population size

I10 Adapt replacement parameters

I11 Adapt stopping condition

I12 Focus on selected users and corresponding role assignments

(e.g. using speciﬁc mutation or crossover operators)

I13 Exclude selected users (Exclude users (and corresponding

role assignments) from the optimization process)

I14 Store a role concept (Transfer an individual from the current

population into the role concept repository)

I15 Insert a role concept into the population (Insert an individual

of the role concept repository into the current population)

by Xin et al. in (Xin et al., 2018)

In the following, an overview of user interac-

tion events relevant in the context of role mining is

provided. These events are given identiﬁers I01 to

I15 and are grouped into four categories, see Ta-

ble 1. The ﬁrst category (I01-I06) contains interac-

tion events leading to a direct manipulation of role

concepts. Roles and the corresponding assignment to

users can be edited, added or removed. A DM can

try to include his or her expert knowledge into the op-

timization process, for example by adding roles that

have proven to be particularly good in the past (I01) or

by removing roles that seem unpromising (I02). How-

ever, it is not guaranteed that these suggestions are

also well-suited in the current role mining scenario. It

should therefore be possible for the optimization al-

gorithm used to reverse the proposed changes later on

in the optimization process, if the expected improve-

ment is not obtained.

The second category (I07-I11) contains adjust-

ments of the parameters of the role mining algorithm

used. Since these indirect interactions strongly de-

pend on the role mining approach used, some exam-

ples for EAs are provide at this point.

Analogous to the approach of Nascimiento, the

DM should be provided with the option to set the

focus of the optimization algorithm to certain areas.

Such events are contained in the third category (I12-

I13). For example, the role mining algorithm could

be prompted to focus particularly on the users of a

certain department of the company if no satisfactory

roles for the users of this department are included

in the proposed role concepts yet (I12). Addition-

ECTA 2023 - 15th International Conference on Evolutionary Computation Theory and Applications

154

ally, in order to reduce the problem size, a DM can

decide to exclude users and the corresponding roles

from further optimization (I13). Since this means that

the roles assigned to the excluded users can no longer

be modiﬁed, which in turn implicitly affects the fur-

ther optimization process, this interaction possibility

should be treated with great foresight and caution.

The fourth category (I14-I15) addresses the pre-

vailing risk of EAs to get stuck in local optima. Indi-

viduals obtained in previous iterations can be stored

and reintegrated into the population of the evolution-

ary algorithm, whenever necessary, thereby avoiding

the necessity of a complete restart of the optimiza-

tion process. In particular, in dynamic optimization,

the ﬁtness landscape is subject to change over time.

Hence, it is possible that some of the stored indi-

viduals may have better ﬁtness values than the indi-

viduals of the current population. Therefore, inject-

ing such stored individuals into the current popula-

tion appears to be a promising approach. This is also

referred to as memory-based evolutionary algorithms

and has been covered in previous publications. A sur-

vey on memory-based evolutionary algorithms is pro-

vided by Branke (Branke, 1999). Branke proposes to

adopt the concept of memory-based evolutionary al-

gorithms to the domain of role mining with user inter-

action. In this scenario, the DM can store interesting

role concepts within a so-called role concept repos-

itory (I14). Hence, the DM is able to analyze and

possibly deploy these role concepts later (I15).

5 INTEGRATION OF

INTERACTION EVENTS INTO

EVOLUTIONARY

ALGORITHMS

One advantage of EAs considering the inclusion of

dynamically occurring interaction events is their iter-

ative procedure. At the beginning of each iteration, it

can be checked whether one or more events are cur-

rently pending. If this is the case, the correspond-

ing event-handling methods can be executed. Fig-

ure 3 shows the alteration of the sequential process of

the addRole-EA for the integration of event-handling

methods.

In addition, it is recommended not to terminate the

execution of the addRole-EA according to (SC1) or

(SC2), but to continue the role mining process, when-

ever computing capacity is available. In this way, the

role concepts can potentially be further improved and

events can be processed close to real-time. Since this

approach is independent of the speciﬁc features of the

Figure 3: Integration of interaction events into addRole-EA.

addRole-EA, it can be transferred to any other evolu-

tionary algorithm.

6 EVALUATION AND

EXPERIMENTS

In this section, the events of the ﬁrst event category

deﬁned in Section 4 are examined in more detail, as

these include direct manipulations of individuals and

therefore impose a particular challenge from an al-

gorithmic point of view. Special focus is given to

interaction event I01, where a DM proposes poten-

tially good roles, as well as I02, where a DM removes

potentially bad roles, as changes resulting from the

other events of this category can be implemented as a

combination of adding and removing roles. For this

purpose, it is ﬁrst shown how established benchmarks

for the RMP can be extended to be suitable for the

simulation of event I01 and I02 and their evaluation.

Subsequently, the two events selected are examined in

different evaluation scenarios.

6.1 Simulation of Events and

Preparation of Benchmarks

In practice, the deployment of role concepts is often

accompanied by experts and consultants. Over time,

these experts acquire extensive knowledge in the area

of role mining and are therefore able to identify roles

that bear the potential to improve the role concepts

encoded by the individuals of the current population

or to accelerate the optimization process in certain sit-

uations. Such roles will be referred to as good roles

in the following. Likewise, they are able to assess

which roles potentially hamper the optimization pro-

cess. These roles will be referred to as bad roles.

It is evident that it is not possible to transfer this

knowledge onto the synthetic evaluation scenarios

corresponding to available benchmark instances for

the RMP. Therefore, other methods must be found

in order to identify good and bad roles in a given

benchmark instance. For the evaluation of the interac-

tion events in this paper, three benchmark instances of

RMPlib, a publicly accessible library for role mining

Interactive Role Mining Including Expert Knowledge into Evolutionary Algorithms

155

Table 2: Benchmark instances selected for evaluation.

Users M Prms. N Roles K

Density ρ

PS 02 50 50 25 0.240

PS 05 100 100 50 0.137

PM 01 500 500 150 0.062

benchmarks, were selected. The benchmark instance

PLAIN small 02 (PS 02) is a rather small data set,

which has a comparatively high density. The bench-

mark instance PLAIN small 05 (PS 05) is slightly

larger, but less densely populated. The benchmark in-

stance PLAIN medium 01 (PM 01) is a medium-size,

low-density data set. The speciﬁcations of the in-

stances selected are displayed in Table 2. At this, ρ

denotes the density of the UPA matrix corresponding

to the respective benchmark instance. Furthermore,

denotes the number of roles that were used to cre-

ate the benchmark and thus serves as upper bound on

the minimum number of roles. For a more detailed de-

scription of the benchmark instances of RMPlib, refer

to (Anderer et al., 2021b).

In order to identify good roles, the addRole-EA

was run 200 times on each benchmark instance. Sub-

sequently, the 20 best role concepts obtained for each

instance were selected. Based on these results, a role

was classiﬁed as a good role, if it was included in

each of the 20 role concepts. From this, a set of 10

good roles, each of which are assigned between 3 and

8 permissions, was obtained for PS 02. For PS 05,

the set of good roles is comprised of 46 good roles,

each of which are assigned between 2 and 11 permis-

sions. For PM 01, 144 good roles could be identiﬁed,

ranging from 3 to 22 permissions each. A role was

classiﬁed a bad role, if it was not included in at least

one of the 20 best role concepts. Even though the

presented procedure was solely applied to the three

benchmark instances selected, it can be applied to the

other benchmark instances of RMPlib as well. At this

point, it must be noted that, in the context of RMP-

lib, the addRole-EA is already used to identify good

and bad roles before the execution of the role mining

process. This does not correspond to practice, where

this knowledge emerges from the growing wealth of

experience of role mining experts.

6.2 Event I01: Addition of Good Roles

In the following, it is examined whether the addition

of the good roles deﬁned in the last section can actu-

ally enhance the optimization process. For this pur-

pose, random roles are selected from the set of good

roles at different points in time during the optimiza-

tion process. Here, each good role selected induces

an instance of interaction event I01, which is handled

Table 3: Parameter values for the evaluation of I01.

PS 02 PS 05 PM 01

Number of events |E| 3; 5 5; 10 20; 30

Simulated at iteration t

= 5, 000; t

= 10, 000; t

= 50, 000

by adding the selected role to all individuals in the

current population using the addRole-Method of the

addRole-EA. Should it occur that multiple instances

of event I01 are pending at the same time, the corre-

sponding good roles are added to the individuals suc-

cessively by repeatedly calling the addRole-method.

To establish comparability, a copy of the population is

created immediately before the good roles are added.

This copy, which is referred to as Pop

, is further op-

timized without including the instances of interaction

event I01. The population in which the instances of

I01 are included is referred to as Pop

I01

The number of good roles, which are transferred

into the optimization process, is based on the bench-

mark instance used. In each case, up to 20% of the

number of roles used to create the benchmark instance

were selected randomly from the set of good roles

and added to the individuals of the current popula-

tion at t

= 5, 000, t

= 10, 000 and t

= 50, 000. Ta-

ble 3 shows the number of instances |E| of interac-

tion event I01 as well as the timing of the event sim-

ulation resulting in the different evaluation scenarios.

The tests for each evaluation scenario were repeated

20 times with different random seeds and run on a

computer with the following speciﬁcations: processor

Intel Core i5-4570S, 2.90 GHz, 16 GB RAM. The pa-

rameters of the addRole-EA were adopted from (An-

derer et al., 2020), see Figure1.

Figures 4 - 6 show the progression of the number

of roles over iterations r

∗

(t) and r

∗

I01

(t), which refer

to the number of roles of the best individual of Pop

resp. Pop

I01

at iteration t, where ﬁve good roles were

added at t

on PS 02, PS 05 and PM 01. Since similar

effects are observed for all other combinations of pa-

rameter values on the selected benchmark instances,

only one representative graph is shown for each in-

stance in the following.

Figure 4: Number of roles over iterations for I01 on PS 02.

ECTA 2023 - 15th International Conference on Evolutionary Computation Theory and Applications

156

Table 4: Evaluation of the addition of |E| good roles.

= 5, 000 t

= 10, 000 t

= 50, 000

∗

(

t) r

∗

I01

(

t) ϕ(E) t

int

(E) r

∗

(

t) r

∗

I01

(

t) ϕ(E) t

int

(E) r

∗

(

t) r

∗

I01

(

t) ϕ(E) t

int

(E)

PS 02 |E| = 3 31.30 30.95 0.58 330 30.85 31.65 0.27 250 30.35 30.40 0.05 10

|E| = 5 33.55 32.05 0.47 280 30.85 30.40 0.30 150 30.10 30.15 0.10 40

PS 05 |E| = 5 50.15 50.45 0.22 90 50.30 50.10 0.04 50 50.55 50.55 -0.05 0

|E| = 10 50.80 50.15 0.21 100 49.95 50.30 0.03 20 49.95 49.95 0.00 0

PM 01 |E| = 20 153.25 153.00 -0.02 0 153.00 152.70 -0.33 0 151.80 151.65 -0.05 0

|E| = 30 152.70 152.55 -0.10 0 152.45 152.25 -0.38 0 151.85 151.65 -0.00 0

Figure 5: Number of roles over iterations for I01 on PS 05.

Figure 6: Number of roles over iterations for I01 on PM 01.

First, it can be seen that the integration of event

instances signiﬁcantly improves the optimization pro-

cess on all benchmark instances. On PS 02 and

PS 05, however, the integration of interaction events

results in a temporary increase in the number of roles

of the individuals in population Pop

I01

. Since the oc-

currence of interaction events is disregarded in the

control population Pop

, the number of roles of its in-

dividuals remains unchanged. To provide a measure

for this effect, a new key ﬁgure, the impact ϕ(E), is

introduced for a set of interaction events E. It is de-

ﬁned as the difference in the number of roles of the

best individual at the time immediately after applica-

tion of the corresponding event-handling method t

and directly before event occurrence t divided by the

total number of event instances:

ϕ(E) :=

r(I

∗

) − r(I

∗

,t)

|E|

(2)

A further interesting key ﬁgure to describe the course

of the optimization process comparing Pop

and

Pop

I01

is the number of iterations t

int

(E) needed until

the ﬁtness of the best individual in Pop

I01

equals the

ﬁtness of the best individual of Pop

for the ﬁrst time

after the occurrence of the interaction event. In case

ϕ(E) > 0, this corresponds to the ﬁrst intersection of

both curves after event occurrence, see Figure 4, and

is obtained as follows:

int

= min

{

t − t

: r(I

∗

I01

,t) = r(I

∗

,t)

}

(3)

where t

corresponds to the iteration, in which the

good roles were added to the chromosomes of the in-

dividuals of Pop

I01

. In case ϕ(E) ≤ 0, t

int

:= 0. Ta-

ble 4 shows the values obtained for t

int

(E), the im-

pact ϕ(E) as well as the number of roles r

∗

(

t) and

∗

I01

(

t) of the best individuals of both populations after

executing

t = 100, 000 iterations of the addRole-EA.

For easier comprehension, the impact ϕ(E) as well as

int

(E) are illustrated in Figure 4.

According to the observations in Figures 4 and 5,

it can be seen that on PS 02 and PS 05 the values of

the impact are mostly positive. This results from the

fact that new roles are added to the individuals by in-

stances of event I01. Even though it seems logical that

an instance of event I01, which corresponds to the ad-

dition of one role, would increase the number of roles

by one, the values for the impact range between -0.38

and 0.58 and are thus signiﬁcantly smaller than one or

even negative. This is caused by the operation prin-

ciple of the addRole-method. As shown in Section 3,

adding a new role by means of the addRole-method

includes the deletion of existing roles that have be-

come obsolete. Since in this test case, good roles were

added, it can be assumed that existing roles became

obsolete in a relatively large number of cases. On

PM 01, this even leads to negative values of ϕ(E),

which corresponds to an immediate improvement of

the number of roles due to the integration of inter-

action events. Additionally, it can be seen that the

simulation of interaction events at later points in time

causes a reduced impact on PS 02 and PS 05. This

may be due to the deﬁnition of good roles used in this

paper. Whenever good roles are added rather at the

end of the optimization process, it is possible that they

are already included in some of the individuals, so

that these individuals are not affected by the addition

Interactive Role Mining Including Expert Knowledge into Evolutionary Algorithms

157

Table 5: Iterations and time (s) for I01 on PS 02.

Number of iterations needed Computation time (s) needed

k |E| = 0 3 5 |E| = 0 3 5

40 7,370 3,190 2,700 154.62 63.91 53.14

38 10,375 5,270 4,390 204.94 99.68 82.37

36 13,645 8,140 7,920 252.76 143.69 136.23

34 20,700 10,850 12,580 344.45 180.15 198.24

32 - 20,340 - - 291.11 -

Table 6: Iterations and time (s) for I01 on PS 05.

k |E| = 0 5 10 |E| = 0 5 10

75 1,350 430 260 15.78 3.21 4.35

70 3,180 1,470 790 35.48 13.88 9.88

65 4,995 2,690 1,580 53.41 25.33 17.55

60 6,630 4,450 3,300 67.82 40.44 33.18

55 9,860 7,150 5,960 93.98 61.65 54.71

of the roles. In all test setups, that result in positive

impact values, the rather small values of t

int

(E) show

that, although the number of roles r

∗

I01

(t) initially in-

creases due to event integration, it is again at least as

good as r

∗

(t) within a few iterations. Thus, the added

roles require a certain amount of time to reveal their

positive effects on the optimization process.

Table 4 further illustrates that the inclusion of In-

stances of I01 in Pop

I01

leads to better results with

respect to the number of roles in 11 out of 18 cases

t = 100, 000 compared to only 5 out of 18 cases, in

which better results were obtained for Pop

However, the potential of interaction event I01 lies

mainly in the short-term effects. In Figures 4 - 6, it

could already be seen that the inclusion of instances

of event I01 leads to a signiﬁcantly faster reduction

of the number of roles of the best individual of the

associated population Pop

I01

compared to that of the

control population Pop

. In order to examine this in

more detail, for each benchmark instance, different

role levels k were speciﬁed, which serve as reference

values to evaluate the short-term performance of the

inclusion of I01. Tables 5 - 7 show the number of iter-

ations and the computation time needed to attain the

respective role level from the occurrence of the inter-

action events at t

. Again, similar effects are obtained

considering t

and t

, so that the corresponding tables

Table 7: Iterations and time (s) needed for I01 on PM 01.

Number of iterations needed Computation time (s) needed

k |E| = 0 20 30 |E| = 0 20 30

420 1,750 350 300 327,32 52,41 54,80

400 3,045 1,020 770 548,52 156,03 126,99

380 4,155 1,640 1,180 722,81 244,01 184,51

360 5,145 2,160 1,590 863,79 313,47 237,29

340 5,970 2,680 2,020 973,36 378,06 289,43

are omitted at this point.

It turns out that in all cases, the addition of good

roles results in the speciﬁed role levels being achieved

in signiﬁcantly fewer iterations as well as in signiﬁ-

cantly less computation time. Furthermore, it seems

that the more good roles are added, the fewer itera-

tions and computation time is needed. In conclusion,

it can therefore be stated that the inclusion of expert

knowledge, by means of interaction event I01, has the

potential to signiﬁcantly accelerate the optimization

process in short-term consideration and leads to bet-

ter results in long-term consideration in many cases.

6.3 Event I02: Deletion of Bad Roles

In this section, analogous to the study of adding good

roles, it is examined how a DM can transfer his or

her expert knowledge into the role mining process by

removing potentially bad roles. However, while the

addRole-EA already provides a method for the ad-

dition of good roles using the addRole-method, the

removal of bad roles is not readily supported due to

the 0-consistency constraint. If bad roles were simply

deleted without further action, users could lose per-

missions essential to performing their tasks. There-

fore, suitable repair methods must be developed to en-

sure compliance with the 0-consistency constraint af-

ter roles were removed. For this purpose, users lack-

ing permissions according to U PA are identiﬁed in a

ﬁrst step. Subsequently, one of the following repair

methods is executed:

R1: Assign all Roles + Unique Roles. In order to

reassign permissions to the users, which were with-

drawn by the removal of bad roles, ﬁrst, it is inves-

tigated whether some of the existing roles in PA can

be assigned to the users without violation of the 0-

consistency constraint to cover some of the required

permissions. If, thereafter, a user is still lacking per-

missions, a new role is created for each of these per-

missions. This corresponds to the idea of the initial-

ization method of the addRole-EA. An advantage of

this method is that the algorithm gains the possibility

to create new roles from scratch. However, many ad-

ditional roles are required to compensate for the dele-

tion of bad roles using this approach.

R2: Assign all Roles + One Role + Unique Roles.

In order to avoid the creation of many new roles, re-

pair method R2 aims at grouping the uncovered per-

missions into one role. Again, the users are assigned

all roles available in PA, that can be assigned to them

without violation of the 0-consistency constraint. If,

thereafter, a user is still lacking permissions, a new

ECTA 2023 - 15th International Conference on Evolutionary Computation Theory and Applications

158

role is created from the remaining uncovered permis-

sions. If this role is not equal to the bad role, which

was previously removed, it is assigned to the users. If

this role equals the bad role, analogous to R1, a new

role is created for each of the uncovered permissions.

An advantage of this method is that the creation of

many new roles can possibly be avoided. An algorith-

mic description on how the UA and PA matrices of an

individual I are adapted to the removal of a bad role

bad

by repair method R2 is provided in Algorithm 1.

If lines 9 - 17 are replaced by lines 13 - 16, one obtains

repair method R1.

Algorithm 1: Repair Method R2.

Input: UPA, UA, PA, r

bad

Output: UA, PA

1 for each role r corresponding to a row of PA do

2 for each user u ∈ U do

3 if r can be assigned to u without

violation of 0-consistency then

4 Assign r to u;

5 end

6 end

7 end

8 Identify uncovered permissions by comparing

UPA and RUPA;

9 Create new role r

new

, which is assigned all

uncovered permissions;

10 if new role does nor equal bad role r

new

6= r

bad

then

11 Add r

new

to individual I using

addRole-method;

12 else

13 for each uncovered permission p do

14 Create new role r, which is assigned p;

15 Add r to individual I using

addRole-method;

16 end

17 end

Evaluation. The evaluation of interaction event I02

was conducted analogously to the evaluation of event

I01. The parameters on which the various evaluation

scenarios are based can therefore also be found in Ta-

ble 3. At the moment of event occurrence, three iden-

tical copies of the current populations are made. In

Pop

respectively Pop

, for each individual, each

role is examined individually and removed if it was

classiﬁed a bad role as described in Section 6.1. This

is repeated until either all roles of the individual are

examined or the maximum number of roles to be

deleted from an individual, as speciﬁed by |E|, is at-

tained. Subsequently, the compliance of the individu-

als with the 0-consistency constraint is restored once

using repair method R1 and once using R2. Again,

Pop

serves as control population in which I02 is dis-

regarded.

Figure 7 shows the progression of the number of

roles over iterations r

∗

(t), r

∗

(t) and r

∗

(t) which

denote the number of roles of the best individual of

Pop

, Pop

and Pop

at iteration t, where up to ﬁve

bad roles were removed at t

on PS 02.

Figure 7: Number of roles over iterations for I02 on PS 02.

It can be seen that event I02, unlike I01, does not

lead to any signiﬁcant improvements in the optimiza-

tion process either by means of repair method R1 or

R2. The same can be observed considering t

and t

on all of the selected benchmark instances and can be

explained by the operation principle of the addRole-

EA. By repeatedly adding new roles and subsequently

deleting obsolete roles using the addRole-method, it

can be assumed that bad roles will be eliminated au-

tomatically. This is also evident when considering the

long-term effects of I02. For this purpose, Table 8

shows the number of roles of the best individual of

the three populations at

t = 100, 000 as well as the

impact of the interaction events ϕ

(E), when using

repair method R1 respectively ϕ

(E) when using R2.

At t

and t

, in only 4 out of 12 cases, the results ob-

tained from event integration obtained better or the

same results for both repair methods compared to the

results obtained for Pop

. In all the other cases, at

least one of the repair methods provided worse results

compared to Pop

. This suggests that the removal of

bad roles at the beginning of the optimization pro-

cess has no signiﬁcant effect on the optimization pro-

cess which is also underlined by the corresponding

low impact values. At t

, however, the integration of

interaction events leads to better results in all cases

independent of whether R1 or R2 is used. This may

be due to the fact that at the time of event occurrence,

the addRole-EA has almost converged on all bench-

mark instances. The removal of bad roles and the

subsequent addition of new, smaller roles causes for

a modiﬁcation of the individuals and possibly creates

new optimization potential. Table 8 shows that R2

leads to better results compared to R1 in 5 out of 6

cases, which is possibly due to the fact that R2 ﬁrst

attempts to create one single role to cover the uncov-

Interactive Role Mining Including Expert Knowledge into Evolutionary Algorithms

159

Table 8: Evaluation of the deletion of |E| bad roles.

= 5, 000 t

= 10, 000 t

= 50, 000

∗

(

t) r

∗

(

t) r

∗

(

t) ϕ

(E) ϕ

(E) r

∗

(

t) r

∗

(

t) r

∗

(

t) ϕ

(E) ϕ

(E) r

∗

(

t) r

∗

(

t) r

∗

(

t) ϕ

(E) ϕ

(E)

PS 02 |E| = 3 30.35 29.60 30.40 -0.03 -0.02 29.50 29.90 29.90 0.07 0.03 30.00 29.80 29.70 1.12 0.43

|E| = 5 30.65 30.60 29.55 0.03 0.04 30.60 30.30 30.60 0.16 0.13 29.15 29.25 29.05 1.23 0.63

PS 05 |E| = 5 50.25 49.90 50.15 0.00 0.00 49.80 49.85 50.35 0.24 0.17 50.45 50.00 49.70 0.93 0.23

|E| = 10 50.15 50.45 50.10 0.03 0.02 50.05 50.05 50.30 0.46 0.36 50.45 49.80 49.70 0.54 0.23

PM 01 |E| = 20 151.85 151.55 151.25 0.00 0.00 151.35 151.10 152.10 0.07 0.03 151.65 150.55 150.50 0.41 0.41

|E| = 30 151.55 151.95 151.90 0.00 0.00 151.30 151.60 151.15 0.13 0.05 151.70 151.30 151.80 0.68 0.64

ered permission after the removal of bad roles, such

that less new roles are created in total.

A further interesting observation is that the

impact values tend to increase, the later the instances

of I02 occur. This is because roles are assigned

rather few permissions at the beginning of the

optimization process due to the initialization method

of the addRole-EA. These can therefore be used

more frequently to cover needed permissions which

were withdrawn from users removing bad roles.

Toward the end of the optimization process, roles

are more likely to be assigned a larger number of

permissions. Consequently, a larger number of new

roles is necessary to cover the needed permissions.

In summary, it can be concluded that, especially

in comparison with interaction event I01, the deletion

of bad roles does not have a major impact on the op-

timization process. However, at later points within

the optimization process, the removal of bad roles has

proven its potential to further improve the proposed

role concepts.

7 CONCLUSION AND FUTURE

WORKS

This paper investigated the integration of expert

knowledge into evolutionary algorithms for role min-

ing. First, an overview of events relevant in the con-

text of role mining, which mostly occur dynamically

during the optimization process and are triggered by

the interaction of a decision maker with role mining

software, was provided. A framework was presented

on how to integrate interaction events in near real-

time into the iterative optimization process of evolu-

tionary algorithms. Additionally, a method designed

to enable the simulation of interaction events using

known role mining benchmarks was presented. Based

on this method, different events and their effects on

the optimization process were evaluated in a series

of experiments. It was demonstrated that the integra-

tion of expert knowledge, especially by adding poten-

tially good roles, leads to better optimization results

and thus better role concepts in many cases. Fur-

thermore, it causes for a signiﬁcant acceleration of

the optimization process. This is of particular im-

portance in large companies, where several thousand

users work together in an ERP system and therefore

long runtimes of the role mining algorithm are to be

expected. Even if the interaction event corresponding

to the deletion of potentially bad roles has a signiﬁ-

cantly lower impact on the optimization process com-

pared to the addition of potentially good roles, it could

be shown that this can still improve the optimization

result, especially if these are deleted rather at a later

point in time within the optimization process.

The deﬁnition of potentially good roles in this pa-

per ensured that these provide a certain quality in the

context of the respective benchmark instance. In prac-

tice, however, it is possible that assumingly good roles

proposed by a decision maker have no or even nega-

tive inﬂuence on the optimization process in the sce-

nario under consideration. It is therefore desirable

to maintain both, individuals that have been modiﬁed

due to the interactions of a decision maker as well as

unmodiﬁed individuals, in the population for a cer-

tain time. However, it has been shown that in many

cases adding good roles initially increases the num-

ber of roles of an individual (positive impact). Due

to the elitist replacement method of the addRole-EA,

these individuals would be inclined not to be trans-

ferred to the following generations. Analogously, in

case of negative impact values, the original, unmod-

iﬁed individuals would potentially not be transferred

to the subsequent generations. One way to address

this is to develop suitable survival strategies that en-

sure that both modiﬁed and unmodiﬁed individuals

survive long enough to unfold their potential in terms

of improving the optimization process. Furthermore,

the cooperation between users and the evolutionary

role mining algorithm could adaptively be improved

by the introduction of a role repository where experts

can store potentially good roles, so that the addRole-

EA can propose them autonomously in future role

ECTA 2023 - 15th International Conference on Evolutionary Computation Theory and Applications

160

mining projects, e.g. via customized mutation and

crossover methods. Since the presented methods for

the inclusion of expert knowledge have speciﬁcally

been designed for use in practice, an application in

industrial use case scenarios would be desirable.

ACKNOWLEDGEMENTS

This study was funded by the German Ministry

of Education and Research under grant number

16KIS1000.

REFERENCES

Anderer, S., Kempter, T., Scheuermann, B., and

Mostaghim, S. (2021a). The dynamic role mining

problem: Role mining in dynamically changing busi-

ness environments. In B

ack, T., Wagner, C., Garibaldi,

J. M., Lam, H. K., Cottrell, M., Merelo, J. J., and War-

wick, K., editors, Proceedings of IJCCI 2021, On-

line Streaming, October 25-27, 2021, pages 37–48.

SCITEPRESS.

Anderer, S., Kreppein, D., Scheuermann, B., and

Mostaghim, S. (2020). The addRole-EA: A new evo-

lutionary algorithm for the role mining problem. In

Guerv

os, J. J. M., Garibaldi, J. M., Wagner, C., B

ack,

T., Madani, K., and Warwick, K., editors, Proceed-

ings of IJCCI 2020, Budapest, Hungary, November 2-

4, 2020, pages 155–166. SCITEPRESS.

Anderer, S., Scheuermann, B., Mostaghim, S., Bauerle, P.,

and Beil, M. (2021b). RMPlib: A library of bench-

marks for the role mining problem. In Lobo, J., Pietro,

R. D., Chowdhury, O., and Hu, H., editors, Proceed-

ings of SACMAT 2021, Virtual Event, Spain, June 16-

18, 2021, pages 3–13. ACM.

Blundo, C. and Cimato, S. (2010). A simple role mining al-

gorithm. In Shin, S. Y., Ossowski, S., Schumacher,

M., Palakal, M. J., and Hung, C.-C., editors, Pro-

ceedings of the 2010 ACM SAC, Sierre, Switzerland,

March 22-26, 2010, pages 1958–1962, New York,

New York. ACM.

Branke, J. (1999). Memory enhanced evolutionary algo-

rithms for changing optimization problems. In Pro-

ceedings of IEEE CEC 1999, Washington, DC, USA

July 6-9, 1999, volume 3, pages 1875–1882. IEEE.

Do Nascimento, H. A. D. (2003). User Hints for Optimi-

sation Processes. PhD thesis, University of Sydney.

Information Technologies.

Du, X. and Chang, X. (2014). Performance of AI algo-

rithms for mining meaningful roles. In Proceedings

of IEEE CEC 2014, Beijing, China, July 6-11, 2014,

pages 2070–2076. IEEE.

Ene, A., Horne, W. G., Milosavljevic, N., Rao, P., Schreiber,

R., and Tarjan, R. E. Fast exact and heuristic methods

for role minimization problems. In Ray, I. and Li,

N., editors, Proceedings of SACMAT 2008, Estes Park,

CO, USA, June 11-13, 2008, pages 1–10, New York,

New York. ACM.

Ferraiolo, D. F. and Kuhn, D. R. (1992). Role-based access

controls. In Proceedings of NCSC 1992, Baltimore,

Maryland, USA, October 13-16, 1992, pages 554 –

563, Baltimore.

Huang, H., Shang, F., Liu, J., and Du, H. (2015). Handling

least privilege problem and role mining in RBAC. In

Journal of Combinatorial Optimization, volume 30,

pages 63–86.

onig, R. and Schneider, S. (2011). Nutzerinteraktion bei

der computergest

utzten generierung von layouts. In

Donath, D. and K

onig, R., editors, Arbeitspapiere Nr.

8 Informatik in der Architektur, Weimar. Bauhaus-

Universit

at Weimar.

Mitra, B., Sural, S., Vaidya, J., and Atluri, V. (2016). A

survey of role mining. In ACM Computing Surveys,

volume 48, pages 1–37, New York, New York. ACM.

Molloy, I. M., Li, N., Li, T., Mao, Z., Wang, Q., and Lobo, J.

(2009). Evaluating role mining algorithms. In Carmi-

nati, B. and Joshi, J. B. D., editors, Proceedings of

ACM SACMAT 2009, Stresa, Italy, June 3-5, 2009,

pages 95–104, New York, New York. ACM.

PwC (2022). PwC’s Global Economic Crime and Fraud

Survey 2022. PricewaterhouseCoopers.

Saenko, I. and Kotenko, I. (2011). Genetic algorithms

for role mining problem. In Cotronis, Y., Danelutto,

M., and Papadopoulos, G. A., editors, Proceedings of

PDP 2011, Ayia Napa, Cyprus, 9-11 February 2011,

pages 646–650, New York, New York. IEEE.

Saenko, I. and Kotenko, I. (2012). Design and perfor-

mance evaluation of improved genetic algorithm for

role mining problem. In Stotzka, R., Schiffers, M.,

and Cotronis, Y., editors, Proceedings of PDP 2012,

Munich, Germany, February 15-17, 2012, pages 269–

274, New York, New York. IEEE.

Saenko, I. and Kotenko, I. (2016a). Reconﬁguration of

RBAC schemes by genetic algorithms. In Badica, C.

and El Fallah Seghrouchni, A. e. a., editors, Intelligent

Distributed Computing X - Proceedings of IDC 2016,

Paris, France, October 10-12 2016, volume 678 of

Studies in Computational Intelligence, pages 89–98,

Cham. Springer.

Saenko, I. and Kotenko, I. (2016b). Using genetic al-

gorithms for design and reconﬁguration of RBAC

schemes. In Proceedings of PrAISeECAI 2016, The

Hague, Netherlands, August 29-30, 2016, pages 1–9,

New York, New York. ACM.

Vaidya, J., Atluri, V., and Guo, Q. (2007). The role min-

ing problem. In Lotz, V. and Thuraisingham, B., edi-

tors, Proceedings of SACMAT 2007, Sophia Antipolis,

France, June 20-22, 2007, pages 175–184, New York,

New York. ACM.

Vaidya, J., Atluri, V., Guo, Q., and Lu, H. (2010a). Role

mining in the presence of noise. In Foresti, S. and

Jajodia, S., editors, Data and Applications Security

and Privacy XXIV - Proceedings of IFIP 2010, Rome,

Italy, June 21-23, 2010., volume 6166 of Lecture

Notes in Computer Science, pages 97–112, Berlin,

Heidelberg. Springer.

Interactive Role Mining Including Expert Knowledge into Evolutionary Algorithms

161

Vaidya, J., Atluri, V., Warner, J., and Guo, Q. (2010b). Role

engineering via prioritized subset enumeration. In

IEEE Transactions on Dependable and Secure Com-

puting, volume 7, pages 300–314.

Xin, B., Chen, L., Chen, J., Ishibuchi, H., Hirota, K., and

Liu, B. (2018). Interactive multiobjective optimiza-

tion: A review of the state-of-the-art. IEEE Access,

6:41256–41279.

ECTA 2023 - 15th International Conference on Evolutionary Computation Theory and Applications

162