Abstract: Modern cyber attack intrusion detection systems apply network flows with artificial labels to build the ability to detect potential threats automatically. Errors，sample insufficiency，and lack of essential features in artificial labeling would severely restrict the system's capability. It is a fatal flaw that the system could not discern attacking samples from benign samples. Most researchers regard the overall performance measurements as the benchmarks for intrusion detection systems while omitting what they are. It was created to warn people about dangerous network attacks. Hence，the article proposed a voting-based framework for an auto cyber intrusion detection system in an imbalanced dataset environment. Based on the trainable voting network，the framework integrated machine learning techniques and deep learning techniques to solve the problem of imbalanced datasets. The article focused on increasing the precision of fatal attack detection without compromising the system's overall performance. The experimental results suggest that the proposed model runs stable and well overall in these different datasets，and the model promotes the detection rate of the minority class effectively.

Key words: intrusion detection, cyber attack recognition, imbalanced sample dataset, deep learning, machine learning