Synergistic TD-TLSTM: Modeling complex temporal dynamics for enhanced anomaly inference in intrusion detection systems

The rapid growth of network traffic and the increasing complexity of cyber-attacks have made intrusion detection systems more challenging. Conventional anomaly-based methods struggle with multi-scale temporal dynamics, irregular inter-arrival patterns, and complex sequential interactions in modern traffic. This paper proposes a Synergistic Tasmanian Devil Optimization–Temporal Long Short-Term Memory (TD-TLSTM) framework to address these issues through three innovations. First, a temporal convolutional architecture with dilated causal convolutions extracts hierarchical features across multiple time scales. Second, a time-aware Temporal LSTM captures irregular time gaps using an adaptive memory decay mechanism, enabling modeling of long-range dependencies in bursty traffic. Third, a Tasmanian Devil Optimization metaheuristic automatically tunes hyperparameters, balancing detection performance and computational efficiency. The framework is evaluated on the UNSW-NB15 dataset with diverse attacks and realistic traffic. Results demonstrate that TD-TLSTM achieves superior performance on the UNSW-NB15 benchmark compared to classical machine learning and existing deep learning baselines, improving macro-averaged F1-score by 8.7% and 12.3% over standard LSTM and Random Forest classifiers, respectively, while reducing training time by 34%. Ablation studies confirm the contribution of each component, and per-class analysis shows improved detection of minority and stealthy attacks, including reconnaissance. Overall, the framework is effective for practical network security systems requiring high accuracy and efficiency.