MULTI-CONVOLUTIONAL BIDIRECTIONAL LSTM ARCHITECTURE FOR HIERARCHICAL FEATURE EXTRACTION IN NETWORK INTRUSION DETECTION

Abstract

The present investigation puts forth a MultiConvolutional Bidirectional LSTM (MC-BiLSTM) model that is both effective and precise for network intrusion detection. In a nutshell, the model consists of parallel 1D convolutional layers and stacked bidirectional LSTMs that together facilitate the learning of hierarchical patterns from sequential network traffic data. With the help of the multiconvolutional front end, local features are captured at different temporal scales, which makes it possible for the system to identify both short-burst anomalies and long-range behavioral patterns. Next, the BiLSTM layers will model the forward and backward dependencies and this will result in better recognition of subtle and evolving attack sequences. An attentionbased pooling layer will bring forward the most informative time steps increasing interpretability and reducing noise simultaneously. The final classification head will be optimized with the use of regularization and imbalance-aware loss functions so that it can manage real-world traffic distributions. The experimental design will facilitate evaluation on the contemporary intrusion datasets like CICIDS and UNSW-NB15. To sum it up, the MC-BiLSTM architecture is a strong, scalable, and context-aware solution for the detection of various cyber threats with increased accuracy and reliability