Paraguay DNA — Machine Learning for Customs Fraud Detection

Paraguay's Dirección Nacional de Aduanas (DNA) established a risk-based selectivity system for import inspections under Decision No. 643/2015, replacing a fully manual inspection-assignment process. DNA partnered with UC Berkeley's Center for Effective Global Action (CEGA), the University of Wisconsin-Madison and Harvard University, funded by the International Growth Centre (IGC), to build and evaluate a machine-learning algorithm for targeting shipments likely to contain undeclared or undervalued goods.

The goal was to replace inspector judgement with a data-driven risk score that could raise the fraud-detection rate of customs inspections while reducing the number of shipments physically inspected.

The ML algorithm was tested in a randomised-controlled trial embedded directly in DNA's live inspection operations, using structured administrative data — declared values, tariff codes, importer history and country of origin — rather than biometric or social data.

The RCT found the ML algorithm achieved a 73% fraud-detection rate versus 18% for human-selected inspections, a fourfold improvement, while inspectors processed 70% fewer shipments to achieve equivalent audit coverage. Extrapolated to all Paraguayan customs ports, researchers estimated the system could recover over USD200 million annually in unpaid customs duties.

This is one of the most rigorously externally evaluated customs ML systems in a lower-middle-income-country context, but no official documentation of ongoing performance monitoring, bias auditing, or parliamentary oversight has been publicly released, leaving accountability opaque beyond the academic evaluation period.