AI-Driven Resilience Testing for Next-Generation Payment Networks: A Digital Twin Framework on the NSF FABRIC Tested

Abstract

The rapid evolution of next-generation payment networks towards greater interconnectivity and intelligence has rendered them indispensable to national economic security. However, this evolution coincides with an emerging paradigm where Artificial Intelligence (AI) is weaponized to power sophisticated, adaptive cyber-physical attacks, exposing a critical gap in existing defensive postures. Current security assessments, reliant on static compliance checks and scripted penetration testing, are fundamentally inadequate for evaluating a system's resilience against these dynamic, AI-augmented threats that exploit the confluence of digital and physical system layers. This research directly addresses this national security challenge by proposing, implementing, and validating a novel AI-driven resilience testing framework for next-generation payment infrastructures. Our core contribution is an integrated Digital Twin environment deployed on the U.S. National Science Foundation's (NSF) FABRIC national-scale programmable testbed. This framework enables high-fidelity, proactive assessment of payment network resilience within a controlled yet realistic experimental ecosystem. Methodologically, the framework constructs a high-fidelity digital replica of a financial exchange network, incorporating accurate topology, protocol emulation (e.g., SWIFT-like messaging), and synthetic transaction flow modeling. To simulate advanced adversaries, we develop automated attack agents using Deep Reinforcement Learning (DRL). These agents are trained to autonomously discover and execute complex, multi-stage attack vectors—such as low-and-slow DDoS and AI-enhanced lateral movement—by interacting with the Digital Twin, with their reward function optimized to maximize systemic disruption or transaction latency. Concurrently, the framework integrates a Multi-Agent System (MAS) to model and evaluate the effectiveness of various elastic defense strategies (e.g., dynamic re-routing, resource scaling) against these AI-powered incursions. Comprehensive experimental evaluation conducted on the NSF FABRIC testbed demonstrates the framework's significant efficacy. In simulated scenarios replicating a tiered financial exchange network, the AI-driven attack agents successfully identified and exploited sophisticated vulnerabilities. Quantitative analysis shows that our framework uncovered 37% more deep-seated and complex vulnerability chains compared to conventional penetration testing tools using predefined scripts. Furthermore, the Digital Twin environment accelerated the validation and comparative analysis of different resilience and recovery strategies by approximately 60%, providing clear, data-driven insights into their performance under duress. In conclusion, this work substantiates that an AI-driven Digital Twin framework, hosted on a national research infrastructure like FABRIC, provides a transformative, proactive, and scalable paradigm for resilience testing. It moves beyond reactive security by enabling the anticipatory evaluation of critical financial infrastructure against the next generation of AI-empowered, adaptive threats. The proposed approach offers a vital empirical platform for researchers and policymakers to develop robust mitigation strategies, thereby contributing directly to the reinforcement of national economic security in an era of increasingly intelligent cyber risks.