SPAM CHALLENGED BY GAME THEORY

Professor Sanjay Chawla, from the School of Information Technologies, a specialist in pattern and data mining and PhD candidate Fei Wang recently published their research results in the journal Machine Learning. Their research was supported by the Capital Markets Cooperative Research Centre (CMCRC). Repetitive game theory is the study of strategic decision making, used across a wide field of behavioural relations including cyber security or financial trading. It is used extensively in financial markets and politics today to analyse and predict decision-making and is now being applied to new areas like computer science and data mining, says Professor Chawla. “It has the potential to be applied to network security situations and in particular to the detection of insurance fraud” he states. “We used game theory because we wanted to apply the concept of adversary and strategic interaction to traditional machine learning approaches to spam classification. “Typical spam filters make more mistakes over time as the spammers work out how to get around the filter” Professor Chawla explains. “An example of this is spammers using misspelt words in the title. This type of adversarial behaviour will produce in a more accurate filter that deteriorates at a much slower rate than current filters would. This means the filter doesn’t need to be upgraded as often reducing the cost, time and disruption associated with upgrading software.” PhD candidate Fei Weng says “Modelling the interaction between a classifier and an adversary as a repeated game theory setting is a far more realistic way of getting training data for the classifier because it allows for cause and effect behaviour to be captured. “We looked for a compromise solution, or equilibrium, where no party wants to deviate from the situation they are in. The classifier is then trained using this equilibrium position. This approach is both more robust and economical that previous methods. We are now investigating ways of integrating sparse learning methods to make the classifier even more robust against adversarial manipulation.” This technique can be applied to many situations wherever there is some motivation for one party to get an advantage over another and has also been used by Fei Wang for analysis of health insurance claims. “We came up with a few models and tested it on real spam data sets. It showed that the model that we built was more robust to changes,” Professor Chawla says. “We’re trying to figure out if there’s a way we can pinpoint different types of attack, like a denial of service. So we’re trying to come up with a unified approach.”