An alternative is to use a method that finds the best combination of features. If you have many features, compared to the size of the dataset, then there is a risk of overfitting. But that can be controlled by using a highly regularized method like Support Vector Machine.

If you need more information please give more details. Are your features binary, scalar, or categorical? How many are there? How large is the training set? What is the intrusion rate in the training set?

Features selected and ML algorithm, which maps input (i.e., features) to output, are tightly linked. I would consider ensembling several ML models. Ideally, they all should be weak learners.

Since all of them have different “understanding”" of the data and thus different feature sets and mapping functions,, they are supposed to make varied uncorrelated predictions. While making the final decision, we take all of these predictions into account and expect the result to be more robust, accurate and less likely biased.

Dear Chris Morris thank you for your immediate answer. I am Using NSL_KDD data set with 125973 instances and 41 features. I am trying to combine 7 (seven classical filter methods). The features are Numeric and Nominal ( I am working by changing from numeric to nominal). My classes are five i.e normal and other four are types of intrusions and the distribution of the training set for classes is Normal -67343 (53.45%), Dos -45927(36.45%), Probe -11656 (9.25%), U2R-52 (0.04%) and R2R - 995 (0.78%). After combining individual feature subsets and form ensemble subset, i plan to use correlation feature selection to select the most relevant feature subset . Valerii Podymov Thank you

There is a challenge to define your training objectives. Accuracy is not the right one, since you can have 99.96% accuracy for U2R-52 by never predicting it. You need to decide what false positive rate is acceptable.