关键词: 鸽群优化算法, 边界过采样, 多头注意力, 双向循环门控单元, 入侵检测

Abstract: In network intrusion detection systems, previous systems often suffered from interference by noisy features during feature extraction, exhibited poor distinction at the boundaries of minority class samples when dealing with imbalanced data, and the detection models tended to miss important temporal information. Such issues compromised the training effectiveness and diminished the detection performance of the models. To overcome these challenges, this paper introduces a hybrid model that amalgamates the pigeon-inspired optimization (PIO) algorithm and the borderline synthetic minority over-sampling technique (SMOTE) with a Transformer-bidirectional gated recurrent unit (BiGRU). The pigeon-inspired optimization algorithm is utilized for automatic feature selection, enhancing the capability of the model to process complex datasets, and mitigating the impact of noise features. The borderline SMOTE is applied to balance the data, with a specific emphasis on minority class samples, thereby enhancing their representation and quality within the balanced dataset. A deep learning model that integrates Transformer and BiGRU is developed for intrusion detection. This integration exploits the Transformer’s ability to capture global dependencies and the BiGRU’s competence in temporal sequence modeling, enabling a more nuanced understanding of the bidirectional contextual relationships in sequence data. Experimental results derived from the NSL-KDD dataset indicate that the model demonstrates commendable detection performance, attaining an accuracy rate of 83.64% and an F1 score of 78.41%, which surpasses the benchmarks set by traditional machine learning models and other deep learning models used for comparison.

Key words: pigeon-inspired optimization algorithm, Borderline SMOTE, multi-head attention, bidirectional gated recurrent unit, intrusion detection