Collective Classification for Spam Filtering - Computational Intelligence in Security for Information Systems - page 23

Table 1. Precision of Boosting, C4.5 and Greedy-Boost

Precision

Class Boosting C4.5 Greedy-Boost

0:Normal 82.6 66.7 99.1

1:Probe 99.0

2:DoS 95.0 99.9 100.0

3:U2R 88.5

4:R2L 93.2

Table 2. Recall of Boosting, C4.5 and Greedy-Boost on KD99

Recall

Class Boosting C4.5 Greedy-Boost

0:Normal 82.0 99.8 100.0

1:Probe 97.1

2:DoS 96.7 89.2 100.0

3:U2R 44.2

4:R2L 71.9

Table 3. Results of the cost

Boosting C4.5 Greedy-Boost

Cost average 0,1467 0,1949 0,0060

Standard deviation 0,5307 0,6067 0,1358

C4.5 and Boosting. Greedy-Boost has more the average low costs (0.0060), this
thanks to its effectiveness on the minority classes, because the minority classes
are the classes whose cost of bad classification is most important.

5 Conclusion and Perspectives

This study provides an ecient design method to generate a robust ensem-
ble system for the intrusion detection problem. The main motivation behind
this method is that the use of aggregation decision for classification ensures the
strengthens of the ensemble system in term of both quantitative and qualitative
sides of members, but also achieves a high diversity degree between members.
The keystone of our proposed solution approach is systematically addressed
through a modification of boosting, specifically, the techniques to update the
weight of examples badly classified. Our theatrical study proves that our ap-
proach greedy-Boost converges quickly. It is experimentally proven that, as a
whole, our algorithm clearly outperform several other algorithms, when evalu-
ated on the KDD99 dataset. We believe that our attempt in this paper is a useful
contribution to the development of ensemble-based intrusion detection systems.