Researchers from Capital Markets Cooperative Research Centre (CMCRC) have developed a spam classifier which uses a model based on repetitive game theory.
Game theory studies strategic interaction between individuals in situations called games. For example, researchers simulated games between an adversary or spammer and the classifier. The classifier learnt from these attacks to predict future attempts by spammers.
Developed by professor of pattern and data mining at the University of Sydney, Sanjay Chawla, CMCRC PhD candidate Fei Wang and former CMCRC PhD student Wei Lu, the classifier is designed to outperform the current model used in email applications.
- Ashes cricket scam targets Australians
- Tax office warns about new email scam
- Perisher Resort stops spambot avalanche
Chawla told Computerworld Australia that the quality of traditional spam classifiers deteriorates over time as spammers work out how to get around the filter.
“For example, you can see in spam emails that the spammer misspelt words such as money. This is because money could be a word that flags the email as spam,” he said.
“By anticipating this adversarial behaviour, it has resulted in a more accurate filter that deteriorated at a much slower rate than a normal spam classifier.”
According to Chawla, this also meant the filter did not need to be upgraded as often, reducing the cost, time and disruption associated with software upgrades.
Wang said that modelling the interaction between a classifier and an adversary as a repeated game theory is a more realistic way of getting training data for the classifier because it allows for cause and effect behaviour to be captured.
“We look for a compromise solution, or equilibrium, where no party wants to deviate from the situation they are in. The spam classifier is then trained using this equilibrium position.”
The research has been published in the Machine Learning Journal.
Follow Hamish Barwick on Twitter: @HamishBarwick