Research on Computer Network Security Posture Prediction Based on BiLSTM

In order to improve the prediction ability of cyber security posture, a time series modeling framework based on BiLSTM is constructed, based on two datasets, NSL-KDD and CIC-IDS2017, fusing 121-dimensional high-dimensional heterogeneous features, and generating 150,000 time series samples using a sliding window mechanism, with the input step size set to 20. The study designs the bi-directional LSTM structure, sets up 64 in each direction units, using ReLU activation with Dropout regular optimization, and completing the training in PyTorch environment by Adam optimizer. The analysis concludes that BiLSTM achieves 93.9% in prediction accuracy with an F1 value of 0.936, which is better than the traditional DNN and unidirectional LSTM models, and has strong generalization and timing modeling capabilities.