A Memory-Efficient Bi-GRU Based Intrusion Detection System with Attention and Probability Calibration for Uncertainty Estimation

This paper proposes a memory-efficient BiGRU-based Intrusion Detection System (IDS) with an Attention mechanism to better learn network traffic patterns and improve attack detection. To make the model more reliable in real-world use, it also includes probability calibration and uncertainty estimation, so the system can indicate how confident it is about each prediction. In addition, focal loss, label smoothing, and oversampling are used to handle class imbalance, while model optimization techniques keep the system lightweight and suitable for resource-limited devices. Experimental results on the CIC-IDS-2018 dataset show that the proposed approach achieves high detection performance while providing more trustworthy predictions for practical network security applications.