An Artificial Intelligence-based System for Detecting Meta Fake Profiles via Gradient Boosting and Multilayer Perception

Abstract

The rise of metaverse platforms has renewed interest in detecting fake profiles, which pose a significant threat to digital ownership and asset transactions within these virtual environments. If digital ownership is not guaranteed, platforms risk missing the point of the metaverse. Current supervised learning techniques for fake profile detection often struggle to maintain acceptable accuracy and interpretability in practice. To address this problem, this study investigates the application of two machine learning models, multilayer perceptron and gradient boosted trees, for detecting fake profiles, with model evaluation performed via two Explainable AI (XAI) techniques, Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP). The algorithms were implemented and evaluated on a dataset of 1244 profiles (1043 real and 201 fake) with 12 attributes. The major finding is that both techniques perform well, but the gradient boosting model achieves a higher accuracy of 99.2% compared with the multilayer perceptron's accuracy of 98.4%. Furthermore, the LIME and SHAP analyses provide insights into the feature importance and decision-making processes of the models. These results suggest that gradient boosting, in conjunction with explainable AI methods, is a more accurate, interpretable, and representative solution for detecting fake profiles in real-world metaverse scenarios, contributing to the development of more secure and reliable digital ownership within these platforms.