Developing an Original Adversarial Robustness Theory for Machine Learning Models in High-Stakes Nigerian Applications

Machine learning models deployed in high-stakes Nigerian applications, including credit scoring, disease diagnosis, and security surveillance, are vulnerable to adversarial inputs that can cause systematic misclassification with serious consequences, yet existing adversarial robustness theory does not account for the distribution shifts and deployment constraints characteristic of Nigerian contexts. This study develops an original adversarial robustness theory for ML models in high-stakes Nigerian applications, extending existing robustness frameworks to accommodate three contextual dimensions absent from prior theory. The theoretical contributions comprise: a contextual threat model that formalises adversarial attack surfaces unique to Nigerian application contexts, including data poisoning through compromised ground-truth labelling in low-oversight data collection pipelines, input perturbation under low-quality imaging conditions, and model extraction through high-volume query fraud; a distribution shift robustness formalism that quantifies model vulnerability to the systematic differences between international training data distributions and Nigerian deployment data distributions; and an original robustness certification method for models deployed under partial compute constraints, based on randomised smoothing adapted for models running on edge hardware. The theoretical framework is validated through empirical application to malaria diagnostic, credit scoring, and face verification models deployed in Nigerian contexts. Robustness evaluations reveal previously unquantified vulnerabilities. The study constitutes an original contribution to adversarial robustness theory applicable to AI deployment in the global south.

Keywords: adversarial robustness, machine learning security, Nigeria, distribution shift, high-stakes AI