TaxRadar: AI-Driven Detection of Unreported Use Tax in E-Commerce

Abstract

Use tax non-compliance in e-commerce constitutes one of the most persistent and under-studied forms of tax revenue leakage in modern fiscal systems. Unlike sales tax, which is collected at point-of-sale by registered merchants, use tax depends on voluntary self-assessment by individual purchasers, a mechanism that consistently produces significant under-remittance. This paper proposes, formalizes, and evaluates TaxRadar, an AI-driven detection framework for identifying unreported use tax obligations in e-commerce transaction streams. The framework integrates a seven-layer data processing and inference architecture, a probabilistic taxability classification model combining fine-tuned gradient-boosted ensembles with deterministic jurisdiction rule constraints, a graph neural network (GNN) module for entity-relationship-based risk propagation, and a composite risk scoring function with formal optimization objective. This paper presents a comparative experimental evaluation on a synthetic benchmark (UseTaxBench) against 13 baselines—spanning rule-based systems, classical ML, and state-of-the-art GNN fraud detectors—suggesting that the proposed framework can improve detection accuracy and reduce false positives relative to traditional rule-based systems on synthetic benchmark data. Privacy-preserving deployment strategies employing federated learning and differential privacy are analyzed. The framework is designed for integration with government tax authority platforms and ERP systems, enabling automated return pre-filling, audit prioritization, and taxpayer notification at scale.