WADCS: Weibull-attentive deep learning model for enhanced cyber security

This work introduces WADCS (Weibull Attentive Deep CyberSecurity), a novel intrusion detection method for DNP3 based smart grid networks that combines distribution aware anomaly modelling with attention driven temporal learning. Unlike existing attention based IDS approaches that only highlight important time steps, WADCS integrates Weibull distribution anomaly scoring with Adaptive Weibull Parameter Estimation (AWPE) to explicitly capture the statistical rarity of cyber events and dynamically adapt to concept drift in evolving network traffic. By aligning low probability Weibull tail events with temporal attention weights, WADCS provides intrinsic interpretability, ensuring that rare but critical anomalies are both detected and explained in real time. Experimental evaluation on the benchmark DNP3 dataset shows that WADCS achieves 94.09% accuracy with a validation loss of 0.07, outperforming conventional GRU and LSTM models while offering a unique combination of distribution aware anomaly detection and temporal attention that is absent in prior IDS solutions. Beyond baseline evaluation, WADCS was rigorously tested under zero-day attack scenarios on multiple datasets, CICIDS2017 and X-IIoTID. Results show that WADCS consistently achieved higher zero-day recall (0.76–0.82) and lower false positive rates (2.9–4.1%) than GNN, FL, Lightweight, and XAI-based IDS, while maintaining competitive efficiency with 3.01 ms latency and 315 samples/sec throughput.