Original Theoretical Investigation of Adversarial Machine Learning Robustness for Intrusion Detection Systems in Nigerian Financial Networks

Adversarial machine learning attacks on intrusion detection systems pose critical security risks to Nigerian financial networks, where AI-based anomaly detection is increasingly deployed, yet the robustness of these systems under adaptive adversarial conditions has not been theoretically or empirically investigated in the Nigerian financial network context. This study conducts an original theoretical investigation of adversarial ML robustness for IDS deployed in Nigerian financial networks. Three original contributions are presented: a formal threat model specifying the attack surface, knowledge assumptions, and perturbation constraints applicable to Nigerian financial network intrusion detection; an adversarial robustness evaluation framework applying projected gradient descent, Carlini and Wagner, and boundary attacks against five representative IDS architectures on a network traffic dataset generated from a simulated Nigerian interbank settlement network; and an original Certified Adversarial Robustness Bound derived from randomised smoothing theory calibrated to the distributional characteristics of Nigerian financial network traffic. The framework was validated against 180,000 simulated normal and attack traffic samples across 12 attack categories. Available adversarial ML literature from financial network IDS contexts identifies evasion attacks exploiting feature space blind spots in flow-based intrusion features as the most technically sophisticated threat vector. The Statistical Learning Theory of Vapnik and the Adversarial Machine Learning Framework by Huang and colleagues provide the theoretical foundation. Findings demonstrate that ensemble IDS architectures with randomised smoothing achieve certified robustness guarantees against perturbations within epsilon equals 0.05 L2-norm constraints. Keywords: adversarial machine learning, intrusion detection, Nigerian financial networks, robustness, cybersecurity.