Adversarial Machine Learning in Digital Payments: A Framework for Detecting and Mitigating Evasion and Poisoning Attacks

Abstract

While Artificial Intelligence (AI) offers unprecedented capabilities for fraud detection and risk assessment in the digital payments ecosystem, the AI models themselves have emerged as a new, high-value attack surface. This paper provides a comprehensive analysis of the emerging threat landscape of adversarial machine learning (AML) in finance, with a specific focus on evasion, data poisoning, and model extraction attacks that can undermine the integrity of payment systems. We argue that traditional cybersecurity controls are insufficient to protect AI systems from these unique threats. To address this gap, we propose a comprehensive AI Resilience Framework for financial services. This framework integrates governance principles from the NIST AI Risk Management Framework (RMF) and MITRE ATLAS, specifies a secure ML-pipeline architecture (MLSecOps), details defense-in-depth mechanisms such as adversarial training, and outlines a robust program for adversarial testing and red teaming. This framework provides a practical, structured roadmap for financial institutions to build secure, robust, and trustworthy AI systems capable of withstanding sophisticated adversarial manipulation.