hide
Language(s):
eng - English
Dates:
2008-06-242008
Publication Status:
Issued
Pages:
52 p.
Publishing info:
Saarbrücken : Universität des Saarlandes
Table of Contents:
In most of the machine learning approaches, it is commonly assumed that the data is independent and identically distributed from a data distribution. However, this is not the actual case in the real world applications. Hence, a further assumption could be done about the type of noise over the data in order to correctly model the real world. However, in some application domains such as spam filtering, intrusion detection, fraud detection etc. this assumption does not hold as there exists an opponent adversary that reacts the filtering process of the classifier and modifies the upcoming data accordingly. Hence, the performance of the classifier degrades rapidly after its deployment with the counter actions of the adversary.
When not assuming the independence of the data generation from the classification, arise a new problem, namely the adversarial learning problem. Now, the classifier should estimate the classification parameters with considering the presence of the opponent adversary and furthermore has to adapt itself to the activities of it.
In order to solve this adversarial learning problem, a two-player game is defined between the classifier and the adversary. Afterward, the game results are resolved for different classifier losses such as adversary-aware and utilitybased classifier, and for different adversarial strategies such as worst-case, goal-based and utility-based. Furthermore, a minimax approximation and a Nikaido-Isado function-based Nash equilibrium calculation algorithm are proposed in order to calculate the resolved game results.
Finally, these two algorithms are applied over a real-life and an artificial data set for different settings and compared with a linear SVM.
Rev. Type:
-
Identifiers:
BibTex Citekey: SönmezMaster 2008
Degree:
Master
