You are on page 1of 5

6.

Experimentation The details of CDEs, DAEs and Normal data items


has already been given in Section 3 and examples
In this section, we describe the method of
have been discussed in Section 5.
evaluation of the proposed algorithm. Firstly, we
describe our dataset. We then calculate various The access pattern data hereby shows that CDEs are
accuracy measures considering different rarely accessed, that too only by a few user roles
parameters as reference. and hence, protection of CDEs from malicious
access is of a greater significance as compared to
6.1 Description of dataset
DAEs and Normal data elements.
This paper is about anomaly detection of user
behaviours. An ideal dataset should be obtained
from a practical system with concrete job functions. 6.2 Cluster Analysis
But in fact, it is very sensitive for almost every
organization or company. When the number of users/user roles exceeds a
given limit, it becomes exceedingly difficult for the
The performance of the algorithm was IDS to keep track of individual user access patterns
analyzed by carrying out several experiments on a and hence detect anomaly. This is the reason that
credit card company dataset adhering to the TPC-C clustering is a better and computationally efficient
benchmark[]. The TPC-C schema is composed of a solution for better performance of IDS. We prefer
mixture of read only and read/write transactions Fuzzy clustering over hard clustering. Fuzzy
that replicate the activities found in complex OLTP clustering (also referred to as soft clustering) is a
application environment. The database schema, form of clustering in which each data point can
data population, transactions, and implementation belong to more than one cluster. In non-fuzzy
rules were designed to broadly represent modern clustering (also known as hard clustering), data is
OLTP systems. We used two audit logs: one for divided into distinct clusters, where each data point
training the model and the second for testing it. The can only belong to exactly one cluster. In fuzzy
training log comprised of normal user transactions clustering, data points can potentially belong to
and testing log consisted of a mixture of normal as multiple clusters. Membership grades are assigned
well as malicious user transactions. Although there to each of the data points(tags). These membership
are unusual records in real dataset, we also inject grades indicate the degree to which data points
some anomalies for detection. The injected belong to each cluster. Thus, points on the edge of
anomalies are set differently with the normal a cluster, with lower membership grades, may be in
behaviour pattern from several aspects. In totality, the cluster to a lesser degree than points in the
about 20,000 transactions were used. In total, center of cluster. When we evaluate various
about 99% of data was non-malicious while less performance measures keeping the number of
than 1% of data was malicious. Fig. 6(a) shows the clusters as a reference parameter, it is observed
distribution of malicious and benign data in the that a particular count for clusters is the most
dataset used: efficient in predicting results.

Fig 6(b) Variation of performance with number of


Fig 6(a) Frequency of data items and their access
clusters
frequency
Fig 6(b) depicts variation in precision, recall, TNR, increase in value of 𝛿1, while the value of Recall
accuracy with change in number of clusters. From decreases with increase in value of 𝛿1.
the graph, we can see that :-
Fig 6(e) shows the variation of Precision, recall, TNR,
 TNR does not vary with the number of clusters, accuracy with 𝛿2. It can be observed from the graph
i.e. TNR is invariant. that the value of Precision, TNR and Accuracy starts
 The precision is always greater than 0.94 and is decreasing when the value of 𝛿2 increases beyond
more or less constant. a certain value. Recall, on the other hand, increases
 Recall reaches optimum value when number of for higher values of 𝛿2.
Fuzzy Clusters is greater than 3.
Fig 6(d) shows the variation of Precision, recall,
 Accuracy also reaches the optimum value
TNR, accuracy with фUT. It can be observed from the
when number of clusters is greater than 3.
graph that the value of Precision first decreases and
6.3 Distances and thresholds then exponentially increases with the increase in
value of фUT. An identical trend is followed by
In section 3.2, we have described Modified Jensen- Accuracy. Somewhat similar trend is followed by
Shannon distance as a measure to calculate TNR except that it does not decrease initially. On
distance between two user vectors of same length. the contrary, the value of Recall decreases with the
In probability theory and statistics, the Jensen– increase in value of фUT.
Shannon divergence is a method of measuring the
similarity between two probability distributions. It Fig 6(f) shows the variation of Precision, recall, TNR,
is also known as information radius (IRad) or total accuracy with фLT. It can be observed from the
divergence to the average. It is based on the graph that the values of all the parameters
Kullback–Leibler divergence, with some notable fluctuate a little but remain more or less constant
(and useful) differences, including that it is with the increase in value of фLT.
symmetric and it is always a finite value. The square
With regards to the dataset we have used, following
root of the Jensen–Shannon divergence is a metric
inferences can be done from the graphs:
often referred to as Jensen-Shannon distance. We
preferred to use modified Jenson-Shannon distance 1. Value of 𝛿1 should be close to 0.65 for
to give weights to data attributes and avoid curse of optimum performance.
dimensionality. The variation of modified Jenson- 2. Value of 𝛿2 should be close to 0.55 for
Shannon distance with Euclidean distance is shown optimum performance.
in the fig 6(g). 3. Value of фUT should be close to 0.59 for
optimum performance.
In section 3.3, we have defined modified Jaccard
4. Value of фLT should be close to 0.2 for optimum
distance to quantitatively measure the similarity
performance.
between two rules. The Jaccard index, also known
as Intersection over Union of the Jaccard similarity
coefficient, is a statistical measure used for
comparing the similarity and diversity of sample 6.4 Comparison with related methods
sets. The Jaccard coefficient measures similarity Table 1 shows the performance measures used for
between finite sample sets, and is defined as the comparison of approaches. Using these
size of the intersection divided by the size of the performance measures, we will compare our
union of the sample sets. The variation of modified approaches with other related works. Our various
Jaccard index with Jaccard index is shown in fig 6(h). approaches are:-

The variation of precision, recall, TNR, accuracy Approach 1. Our approach using modified Jenson-
with the various thresholds, namely 𝛿1, 𝛿2, фUT , фLT Shanon distance and modified Jaccard index.
that were defined in section 3 is shown in the
Approach 2. Using unmodified Jaccard index with
following figures:
Jenson-Shanon distance.
Fig 6(c) shows the variation of Precision, recall, TNR,
Approach 3. Using Euclidean distance with
accuracy with 𝛿1. It can be observed from the graph
unmodified Jaccard index.
that Precision, TNR and Accuracy increase with the
S.No. PERFORMANCE FORMULA
MEASURE
1 TNR TN
TN + FP
2 Precision TP
TP + FP
3 Accuracy TP + TN
TN + FP + TP + FN
4 F1 Score 2 ∗ Precision ∗ Recall
Precision + Recall
5 PPV TP
TP + FP
6 ACC TP + TN
TP + TN + FP + FN
7 NPV TN
TN + FN
8 FDR FP
FP + TP
9 FOR FN
TN + FN
10 BM TPR + TNR – 1
11 FPR FP
FP + TN
12 FNR FN
FN + TP
13 MK PPV + NPV – 1
14 MCC TP × TN − FP × FN
√(TP + FP)(TP + FN)(TN + FP)(TN + FN)
Table 1 ( Performance Measures)
In table 2 we have compared the three approaches If we compare Approach 1 with Approach 3, we
with each other. observe that:-
Sensitivity Approach 1 Approach 2 Approach 3  TNR and precision of Approach 1 is a lot better
Measures
than the TNR and precision for Approach 3
PPV 0.96 0.73 0.74
 .It has also got better accuracy as compared to
TPR 0.81 0.95 1.00
Approach 3.
ACC 0.89 0.80 0.83
 Approach 1 also has a much lower FPR and FDR
F1 Score 0.88 0.83 0.85 score as compared to Approach 3.
NPV 0.83 0.93 1.00  Amongst other performance measures, MK
FDR 0.04 0.27 0.26 and MCC values of Approach 1 are also slightly
FOR 0.17 0.07 0.00 better than that of Approach 3.
BM 0.77 0.60 0.65  Approach 3, on the other hand has got better
FPR 0.03 0.34 0.34 TPR, NPV and FOR measures as compared to
TNR 0.96 0.65 0.65 Approach 1. In fact, it has the best values for
FNR 0.19 0.05 0.00 these parameters in the entire table.
MK 0.79 0.66 0.74  Also, both Approach 1 and Approach 3 have got
MCC 0.78 0.63 0.70 somewhat similar F1 score.

Table 2 (Comparison of our approaches) In the measures like TNR and precision,where
Approach 1 has one of the best score in the entire
From the table, following observations can be
table, Approach 3 performs rather poorly. Also,
made:-
Approach 3 lags far behind in measures like FPR and
If we compare Approach 1 with Approach 2, we can FDR score. On the other hand, in the measures in
observe that:- which Approach 3 performs better than Approach
1, Approach 1 is also performing quiet nicely. For
 TNR and FPR of Approach 1 is a lot better than example, in case of NPV, both the approaches have
the TNR and precision for Approach 2. good scores, with Approach 3 performing better.
 Approach 1 has also got better accuracy as similar trends are observed in case of all other
compared to Approach 2. measures except FNR, where Approach 3 has is far
 Approach 1 has a much lower FPR and FDR superior. Considering all the above scenario, we can
score as compared to Approach 2. say that the overall even though Approach 3 has the
 Amongst other performance measures, MK best values for some performance measures, its
and MCC values of approach 1 are also better poor performance in other measures are clearly a
than that of Approach 2. disadvantage due to which Approach 1 is better
 Approach 2, on the other hand has got better than Approach 3.
TPR, NPV and FOR measures as compared to
Approach 1.
 Both Approach 1 and Approach 2 have got Table 3 shows a comparison of our approaches with
somewhat similar F1 score. various other related works. If we compare our
approach with other related approaches, we
In the measures like FPR and TNR where Approach
observe that:-
1 has good performnce, Approach 2 performs
rather poorly. However, in measures like TPR and  In comparison to HU Panda, our approach
NPV, where Approach 2 performs better, Approach works better with respect to all the
1 also has good performance. For example, both performance measures considered for the
Approach 1 and Approach 2 have similar NPV scores purpose of comparison.
with Approach 2 performing slightly better As  In comparison to the work of Mostafa et al. our
Approach 1 performs far better than Approach 2 in approach performs better with respect to all
most of the measures, we can conclude that the the performance measures that are considered
overall performance of Approach 1 is better than for comparison.
Approach 2.
Sensitivity Approach Approach Approach HU Panda Hashemi Mostafa Mina Majumdar EliSa UP Rao et
Measures 1 2 3 et al. et al. Sohrabi et al. (2006) Bertino al.(2016)
et al. et al.
PPV 0.96 0.73 0.74 0.88 0.97 0.94 0.93 0.88 0.94 0.61

TPR 0.81 0.95 1.00 0.73 0.71 0.75 0.66 0.70 0.91 0.70

ACC 0.89 0.80 0.83 0.81 0.84 0.85 0.80 0.80 0.93 0.64

F1 Score 0.88 0.83 0.85 0.79 0.82 0.83 0.77 0.78 0.92 0.65

NPV 0.83 0.93 1.00 0.77 0.77 0.79 0.73 0.75 0.91 0.68

FDR 0.04 0.27 0.26 0.12 0.03 0.06 0.07 0.13 0.06 0.39

FOR 0.17 0.07 0.00 0.23 0.23 0.21 0.27 0.25 0.09 0.32

BM 0.77 0.60 0.65 0.63 0.69 0.70 0.60 0.60 0.85 0.35

FPR 0.03 0.34 0.34 0.10 0.02 0.05 0.05 0.10 0.06 0.45

TNR 0.96 0.65 0.65 0.90 0.98 0.95 0.94 0.90 0.94 0.65

FNR 0.19 0.05 0.00 0.28 0.29 0.25 0.35 0.30 0.09 0.30

MK 0.79 0.66 0.74 0.65 0.74 0.73 0.66 0.63 0.85 0.29

MCC 0.78 0.63 0.70 0.63 0.72 0.71 0.63 0.61 0.85 0.29

Table 3 (Comparison of our approaches with related works)

 In comparison to the work of Hashemi et al. measures as well, better TNR and precision
even though our approach scores just a little scores can easily cover up lower recall values.
less in measures like TNR and precision, it 7. Analysis and Conclusion
scores a lot better with respect to rest of the
In this paper we have tried to detect malicious
performance measures.
transactions keeping in mind that certain data
 If we consider the work of Mina Sohrabi et al.
elements hold more critical information. We also
our approach performs better with respect to
take into consideration user behaviour pattern in
all the performance measures that are present
this approach. A user regularly behaving as a
in the table.
normal user will be gradually improving his
 In comparison to the work of Majumdar et al.
suspicion score. We then analyse the approach
our approach performs better with respect to
w.r.t different parameters by conducting
all the performance measures that we have
experiments. Finally, we conclude that the
considered for the purpose of comparison.
approach works efficiently in determining the
 With comparison to the work of UP Rao et al.
nature of a transaction.
our approach performs better in context to all
the measures that are considered in the table
for comparison.
 In comparison to the work of Elisa Bertino, our
approach gives better TNR and precision
scores. It also gives comparatively better FDR
and FPR scores. In other measures, except TPR
and recall, both approaches have somewhat
similar score. Since our work is mostly related
to finding Critical Data Items in a dataset,
higher TNR and precision scores are more
desirable as compared to other performance
measures. Since our approach performs quiet
well with respect to other performance

You might also like