TaxFL: Federated Learning and Federated Graph Intelligence for Cross-Border Tax and AML Compliance

Cross-border tax non-compliance and money laundering exploit the same structural weakness: the data needed to identify the beneficial owner of a suspicious structure is fragmented across jurisdictions and institutions. Traditional exchange-of-information (EOI) mechanisms are essential for targeted cases but cannot support the high-volume, pattern-based risk analytics modern compliance requires. This working paper presents TaxFL, a privacy-preserving reference architecture and governance blueprint that lets tax authorities and financial intelligence units collaboratively train and evaluate risk models without sharing raw taxpayer or transaction data — only model updates cross any boundary. TaxFL integrates cross-silo federated learning, federated graph neural networks (GNNs) for ownership and transaction networks, and layered privacy-enhancing technologies (secure aggregation, differential privacy, and optional confidential computing), wrapped in a conservative "belt-and-suspenders" legal blueprint viable even under strict interpretations of international data-transfer rules (e.g., Schrems II, FATF R40). The contribution is deliberately integrative and institutional rather than a new learning algorithm. Rather than claiming that federation is universally beneficial, the paper maps its utility frontier. On a non-IID graph benchmark (Cora) evaluated under a single held-out global test split, federated training — strongest with SCAFFOLD — recovers the discriminative signal that label-skewed local training loses and matches a centralized upper bound, correcting an earlier evaluation artifact. On synthetic, officially-calibrated tax/AML data, the gains are conditional and honestly reported: federation helps at scale (5–10 jurisdictions) and through graph structure over linear models, and SCAFFOLD removes the negative transfer that naive averaging causes, while aggregate gains over an already-strong single silo are modest. The use of synthetic data is intentional — it mirrors the operational constraint that raw taxpayer data cannot cross jurisdictional boundaries and enables fully reproducible peer review. The paper outlines priority use cases (cross-border beneficial ownership, VAT/GST carousel fraud, transfer pricing, crypto-asset risk scoring, and FIU typology sharing), a realistic six-month pilot design for a small multi-jurisdiction consortium starting from public/synthetic data, and an evaluation framework covering detection performance, privacy assurances, operational overhead, and institutional trust. TaxFL is a low-risk complement to existing EOI processes and a deployment-ready blueprint whose core thesis still requires a real-data pilot for validation. Open implementation: https://github.com/pafrantz/TaxFL. We welcome academic and institutional collaborators for rigorous empirical validation on real-world tax data under appropriate safeguards.