Export Reviews, Discussions, Author Feedback and Meta-Reviews

First provide a summary of the paper, and then address the following criteria: Quality, clarity, originality and significance. The paper formalizes the problem of feature cross-substitution in adversarial classification and presents some approaches to solving it (one heuristic and one exact based on mixed-integer linear programming). Previous work in adversarial classification have considered simpler models for how an adversary would try to fool a classifier by replacing values of features (basically, simple distance functions) and assume that the adversary would try to minimize the distance to the real features why fooling the classifier. The paper points out that this approach has a serious shortcoming when feature selection is done. The authors then suggest that it makes more sense to consider equivalent classes of features as basically being the same features and treating them transparently in feature selection and classifier building. They show results of their two approaches (heuristic and exact) on three datasets, outperforming previous work.