A Hybrid Generative AI Framework for Real-Time E-Commerce Fraud Detection: Comparative Performance and Ethical Analysis

Abstract

The rapid growth of e-commerce has made real-time fraud detection a critical challenge. Traditional
rule-based systems fail to adapt to evolving fraud tactics, motivating the adoption of generative artificial intelligence
(AI) models. This paper presents a web-based fraud detection system built on Django that implements and
comparatively evaluates three Generative AI architectures: a Generative Adversarial Network (GAN), a Variational
Autoencoder (VAE), and a novel Hybrid GAN-VAE model. The GAN generates synthetic fraudulent transactions to
address severe class imbalance; the VAE learns latnt representations of normal transaction patterns and flags
anomalies through reconstruction error thresholding; and the Hybrid model combines both strengths, feeding
augmented features into a Random Forest classifier. Experimental results on an e-commerce transaction dataset
demonstrate that the Hybrid GAN-VAE achieves the highest accuracy of 92%, precision of 92%, recall of 90%, and
F1-score of 91%, outperforming standalone GAN (89% accuracy) and VAE (86% accuracy) models, as illustrated in
Fig. 1. The system further integrates the Google Gemini API for AI-powered dataset insight generation. Ethical
considerations including demographic bias analysis, false positive rates across regional groups, and data privacy
compliance are systematically evaluated (Fig. 4). The proposed platform provides a practical, interpretable, and
ethically responsible framework for real-time e-commerce fraud detection.