This paper presents a comprehensive technical investigation into Adversarial Machine Learning (AML) vulnerabilities within US banking fraud detection systems and proposes a resilient defense architecture. We first quantify the catastrophic fragility of conventionally trained Deep Learning models, demonstrating that state-of-the-art Graph Neural Networks (GNNs) are compromised by Projected Gradient Descent (PGD) evasion attacks, achieving an Attack Success Rate (ASR) up to 87.5% . To counter this, we propose the Adaptive Adversarial Defense (AAD) Pipeline, an operational framework integrating continuous threat simulation and Online Adaptive Adversarial Training (OAAT). Empirical results show PGD-AT significantly boosts model resilience, reducing the GNN's ASR to 32.0% and delivering a 52.3% reduction in expected annual fraud loss. The AAD Pipeline provides an economically justifiable blueprint for banks to achieve measurable robustness, addressing the critical need for trustworthy and resilient AI in national financial security.