Detecting Fake Products in African E-Commerce Markets: An Artificial Intelligence Framework Using a Tanzania-Informed Simulation

Abstract

The rapid expansion of e-commerce in Africa has increased consumer access to goods while also creating opportunities for counterfeit and misleading product listings. This study proposes a multi-signal artificial intelligence framework for counterfeit-product risk detection and evaluates it through a reproducible Tanzania-informed simulation rather than through direct observation of Tanzanian marketplace transactions. A synthetic dataset of 1,000 product listings was generated with a 70:30 genuine-to-counterfeit class ratio. Five marketplace signals were modelled: price deviation, seller reputation, image similarity, review sentiment, and description quality. The revised simulation protocol specifies the class-conditional distributions, contamination assumptions, random seed, train-test split, cross-validation strategy, hyperparameter search, and evaluation metrics. Four supervised classifiers - logistic regression, support vector machine, random forest, and XGBoost - were compared. No raw product images were used to train a convolutional neural network; instead, image similarity was treated as a scalar signal that could in practice be produced by an upstream visual-comparison component. On a stratified 20% holdout set, logistic regression achieved 0.920 accuracy, 0.867 precision, 0.867 recall, 0.867 F1-score, and 0.950 ROC-AUC. Repeated stratified five-fold cross-validation produced mean accuracies of approximately 0.889-0.893 across the four models. The findings therefore support the technical feasibility of multi-signal screening under controlled simulated conditions, not operational validation in Tanzanian e-commerce. The paper concludes that future deployment should be validated with real African marketplace data, raw product imagery, network-level seller relationships, and explainable decision support for human moderators