Enhancing Cyber-Physical Resilience in Critical Infrastructure: A Low-Latency, AI-Driven Anomaly Detection Framework for the GERD

The security of Cyber-Physical Systems (CPS) within Critical National Infrastructure (CNI) is of paramount importance. High-value assets such as the Grand Ethiopian Renaissance Dam (GERD) are susceptible to sophisticated False Data Injection (FDI) attacks that can compromise state estimation, leading to catastrophic failures. While Deep Learning has emerged as a powerful defense, existing state-of-the-art Transformer models often prioritize classification accuracy at the expense of computational latency, shrinking the critical window available for mitigation. This paper addresses this trade-off by proposing a novel, modular resilience framework centered around ResiFormer, a bespoke Transformer-based architecture optimized for high-frequency industrial time-series. Unlike traditional Autoencoders or computationally heavy adversarial models, ResiFormer leverages a streamlined Multi-Head Self-Attention mechanism to capture long-range temporal dependencies and complex inter-sensor correlations with minimal inference overhead. We further introduce a novel Latency-Weighted F1-Score to explicitly penalize delayed detections. The efficacy of the framework is validated on a high-fidelity synthetic GERD dataset and the widely adopted real-world SWaT testbed. Benchmarking against the state-of-the-art TranAD model, ResiFormer demonstrates statistically significant superiority, achieving an F1- Score of 0.946 while reducing detection latency by 20% and False Negative rates by over 25%. By delivering high-fidelity alerts within a sub-second timeframe, this work provides a scalable, resilience-aware blueprint for securing next-generation critical infrastructure against process-aware threats.