This paper explores the application of feature selection in intrusion detection using federated learning, aimed at compressing training data features. It addresses the challenges posed by high-dimensional and redundant features in deep learning models, exacerbated in federated learning settings. Leveraging the gain-based importance of decision trees and entropy-based information gain, a collaborative selection algorithm is devised to reduce feature dimensionality. Contributions include a selection methodology, an intrusion detection system employing federated learning, process optimization, and achieving 97.81% accuracy on a subset of the CIDIDS2017 dataset, with training time reduced to one-third.