Nowadays, we see more and more cyber-attacks on major Internet sites and enterprise networks. Intrusion Detection System (IDS) is a critical component of such infrastructure defense mechanism. IDS monitors and analyzes networks' activities for potential intrusions and security attacks. Machine-learning (ML) models have been well accepted for signature-based IDSs due to their learn ability and flexibility. However, the performance of existing IDSs does not seem to be satisfactory due to the rapid evolution of sophisticated cyber threats in recent decades. Moreover, the volumes of data to be analyzed are beyond the ability of commonly used computer software and hardware tools. They are not only large in scale but fast in/out in terms of velocity. In big data IDS, the one must find an efficient way to reduce the size of data dimensions and volumes. In this paper, we propose novel feature selection methods, namely, RF-FSR (Random Forest-Forward Selection Ranking) and RF-BER (Random Forest-Backward Elimination Ranking). The features selected by the proposed methods were tested and compared with three of the most well-known feature sets in the IDS literature. The experimental results showed that the selected features by the proposed methods effectively improved their detection rate and false-positive rate, achieving 99.8% and 0.001% on well-known KDD-99 dataset, respectively.