Developing an Original Theoretical Model of Adversarial Machine Learning Threats to Nigerian Banking AI Systems

Artificial intelligence systems deployed in Nigerian banking for credit scoring, fraud detection, and customer service are vulnerable to adversarial machine learning attacks, and developing an original theoretical model of these threats in the Nigerian banking AI context represents a critical contribution to both AI security and financial cybersecurity theory. This study developed an original theoretical model of adversarial ML threats to Nigerian banking AI systems, drawing on adversarial ML theory, banking AI deployment analysis, and empirical threat modelling. A theoretical model building methodology was employed: systematic review of adversarial ML literature (74 publications from 2019 to 2024), empirical analysis of AI system deployments at 12 Nigerian commercial banks through technical interviews with AI development teams, and threat model design and expert validation. The empirical analysis confirmed that credit scoring models (deployed in 83.3 percent of banks), fraud detection classifiers (75.0 percent), and biometric authentication models (41.7 percent) were the primary attack surfaces. The original Adversarial Threat Model for Nigerian Banking AI proposes four attack categories with Nigerian-specific attack scenarios: evasion attacks on fraud detection models using local payment pattern knowledge, poisoning attacks exploiting training data acquisition processes in poorly secured MLOps pipelines, model extraction attacks targeting credit scoring model parameters, and inference attacks targeting customer financial data from model outputs. Defensive recommendations for each attack category incorporate Nigerian banking operational constraints. Expert review by 16 adversarial ML and banking AI security specialists confirmed the model's original theoretical contribution and practical relevance.