Static prediction games for adversarial learning problems

摘要: The standard assumption of identically distributed training and test data is violated when the are generated in response to presence a predictive model. This becomes apparent, for example, context email spam filtering. Here, service providers employ filters, senders engineer campaign templates achieve high rate successful deliveries despite filters. We model interaction between learner generator as static game which cost functions not necessarily antagonistic. identify conditions under this prediction has unique Nash equilibrium derive algorithms that find equilibrial two instances, logistic regression support vector machine, empirically explore their properties case study on