Research on Intrusion Detection Dataset Construction and Multi-Model Evaluation Method Based on a Unified Framework

The rapid development of cyberspace has led to the complexity of traffic structures and the diversification of attack methods. Traditional intrusion detection methods based on features or rules are increasingly difficult to cope with rapidly evolving threats. In recent years, the introduction of deep learning and Transformer architecture has provided new research directions for Intrusion Detection Systems (IDS). However, current research generally suffers from issues such as inconsistent experimental processes, non-uniform dataset processing standards, and difficult-to-reproduce model evaluations, which seriously hinder academic comparison and engineering implementation. This paper proposes a unified experimental framework for IDS research, which systematically integrates dataset construction, feature preprocessing, class imbalance handling, model training, and performance evaluation to achieve full-process standardization from data to results. The framework supports multiple data sources, multiple imbalance ratios, and multiple model structures, improving experimental comparability and reproducibility while providing a flexible extensibility foundation for subsequent research. Research results show that the framework can significantly enhance experimental consistency and model evaluation efficiency, offering a systematic methodology for model selection and deployment in the IDS field.