The rapid growth of network trac and the high dimensionality of intrusion detection datasets pose major challenges for ecient and accurate intrusion detection. Traditional feature selection methods often rely on heuristic or statistical measures that fail to guarantee representative coverage of the underlying feature space. This paper proposes a principled feature selection framework grounded in submodular optimization. Specically, we formulate the feature selection problem as a submodular set cover task and introduce the Coreset Greedy-Based Feature Selection (CGFS) algorithm to construct compact yet highly representative feature subsets. The proposed CGFS algorithm leverages the facility location function to quantify feature representativeness and applies a greedy selection process that achieves a provable logarithmic approximation bound to the optimal subset. To ensure scalability, we incorporate several eciency-oriented strategies to mitigate the computational cost of large-scale similarity evaluations inherent in the submodular set cover formulation, including pre-normalized inner-product similarity computation, approximate nearest neighbor search via FAISS, and a lazy greedy selection mechanism. Experimental results on ve benchmark IDS datasets demonstrate that CGFS signicantly reduces the number of features while preserving or even improving detection performance, oering a strong balance between accuracy, generalizability, and computational eciency. The source code is available on GitHub.