Filter Feature Selection In Adversarial Environment

Feature selection is a kind of method which lowers the dimension of data in machine learning.The effect of feature selection is that it can remove redundant features from dataset and reduce the computational time.Filter feature selection is one kind of feature selection methods.The traditional one concentrates on selecting features which are benefit for classification,but the problem of security is rarely considered.According to recent research,data subsets generated by traditional feature selection methods is not safe that adversaries can easily evade the detection of classifiers.To overcome this problem,existing approach FAFS proposed adversarial filter feature selection model,which adopts m RMR as generalization ability and averaged distance between every malicious sample to its nearest legitimate samples as security.Then the feature is evaluated by the weighted summation of generalization ability and security.FAFS makes selected features be beneficial to provide effective information for classification and be difficult in modification by adversaries.However,FAFS is not accurate when measuring features with outliers and not suitable to select discrete features,especially Boolean features.In this paper,we proposed a new adversarial filter feature selection methods DAFFS.SU is selected as the generalization ability term,which measures how much information that the feature can provide for classification;at the same time,EM distance,a measurement based on difference between distribution of two classes,is adopted to evaluate security of the feature.Comparing with FAFS,DAFFS has the advantage of less affected by outliers and it is more precise in calculating discrete features.DAFFS can select feature with sufficient generalization ability on the data without attack and with better security on the data under attack,which provides a dependable feature selection method for researches or applications of machine learning problems.DAFFS is suitable for security-sensitive machine learning systems applications like spam filtering or malware detecting.