Learning algorithms for applications involving attacks
Ramarathnam Venkatesan (Venkie), Microsoft Research
Abstract: Learning algorithms have enjoyed a wide use and it has slowly become important to consider scenarios with users who game or attack the system, as well as situations where errors in training sets have correlated errors for which no good model is available. In this vein, we first consider the problem of learning SVMs with adversarial flips of a bounded number of labels, and we present an algorithm and show that is possible to correct errors, and learn a classifier that performs well with respect to the original data distribution. Such models can be useful in applications like anti-spam, document classification, etc. Next we consider the development of learning algorithms based on constructing generative models for frequent patterns in the data. The main advantage of such algorithms is the classifier operates directly in the pattern space allowing for providing reasons/evidence for the classifier decision. We discuss applications to prioritization of attack scenarios based on access control and applications where the security expert seeks to interpret the output of a classifier.
Joint work with Srivatsan Laxman and Prasad Naldurg (Microsoft Research, India)