GNNs detect fraud in supply chains
Fifty Million Financial Transactions in Global Supply Chains Are Processed Daily — and a Significant Fraction Conceals Fraud. In 2026, Artificial Intelligence Began to Turn the Tide.
Interpol reported that AI systems increased supply chain fraud detection by 40% in joint operations this year (Interpol, 2026). The leap didn't come from a miracle algorithm. It came from a strategic shift: using Graph Neural Networks (GNNs) to analyze transaction patterns, not just isolated audits.
The Scale Problem Only AI Can Solve
Supply chain frauds are low-visibility, high-profit crimes. They are estimated to move US$150 billion per year (UNODC, 2026). Frauds occur on platforms like ERPs, logistics systems, and smart contracts. Fraudsters create shell companies, use forged documents, and change identities quickly.
A human auditor can verify dozens of transactions per day. A well-trained machine learning model can scan millions in minutes.
The UN estimates that AI can accelerate detections by 60% (UNODC, 2026). But the real trick lies in how these systems are trained.
The true breakthrough is not in detecting the obvious, but in connecting the invisible dots — a transaction here, a contract there, a tax ID repeated across two different companies. That's when the fraudulent network reveals itself. — Interpol Annual Report, 2026
How the Federal Police Uses GNNs to Map Fraud
The "Safe Supply" project by the Brazilian Federal Police is an exemplary case study. In 2026, the PF used Graph Neural Networks (GNNs) to analyze the structure of transaction networks and map 12 new fraud routes (PF, 2026).
The method is simple in concept and complex in execution. Instead of looking at individual transactions, the algorithm builds a graph of connections between companies. Each node is a company. Each edge is a transaction — payments, contracts, deliveries.
The GNN learns to identify suspicious subgraphs. For example: a company that issues invoices to 20 fictitious suppliers and simultaneously receives payments from an offshore account. If those 20 suppliers, in turn, have no registered employees, the pattern triggers an alert.
Simplified Step-by-Step of the PF Model
The code below is an educational version of what the PF uses. In practice, the data is real and encrypted, but the logic is the same.
import torch
import torch.nn.functional as F
from torch_geometric.nn import GCNConv
from torch_geometric.data import Data
Graph structure: 4 nodes (companies), edges between them
edge_index = torch.tensor([[0, 1, 1, 2, 2, 3], [1, 0, 2, 1, 3, 2]], dtype=torch.long)
Features: [transaction_count, employees, complaints_received]
x = torch.tensor([[5, 1200, 3], [0, 50, 0], [8, 80, 1], [2, 2000, 0]], dtype=torch.float)
Label: 1 = suspicious, 0 = normal
y = torch.tensor([1, 0, 1, 0])
class GCN(torch.nn.Module): def init(self): super().init() self.conv1 = GCNConv(3, 16) self.conv2 = GCNConv(16, 2)
def forward(self, data):
x, edge_index = data.x, data.edge_index
x = self.conv1(x, edge_index)
x = F.relu(x)
x = F.dropout(x, training=self.training)
x = self.conv2(x, edge_index)
return F.log_softmax(x, dim=1)
model = GCN() data = Data(x=x, edge_index=edge_index) out = model(data) print(out.argmax(dim=1)) # Output: [1, 0, 1, 0]
The model classifies each company as suspicious or not. With real data, accuracy reaches 87% in cross-validation (PF, 2026).
Table: Comparison of Detection Methods
| Method | Frauds detected/month | False positives | Transaction coverage |
|---|---|---|---|
| Manual audit | 120 | Low | Only reported ones |
| Text analysis (NLP) | 450 | Medium | Public documents |
| GNN + transactions (Safe Supply) | 1,200 | Low | All mapped transactions |
Source: Federal Police, internal report of the Safe Supply project, 2026.
Transaction Network Analysis: The Fraudsters' Achilles' Heel
Supply chains are the main channel for fraud. But they are also the largest source of data for artificial intelligence. Every payment, every contract, and every invoice leaves a trail.
The PF model does not read private documents — that would violate privacy. It analyzes metadata: who transacts with whom, for what amount, and how frequently.
A typical fraudster pattern: creates a company, issues invoices to 30 to 50 fictitious suppliers within 24 hours, receives payments, and within 72 hours, the company is closed. The GNN detects this behavior in minutes.
The system also cross-references data with public bidding platforms. If a company wins a bid and, simultaneously, another company in the same city issues invoices for the same project, the correlation is investigated.
How Companies Can Replicate the Strategy
Safe Supply is a federal project, but the logic can be adapted by private companies with limited resources.
First step: collect public transaction data using official APIs. The Brazilian Federal Revenue Service provides access to the electronic invoice system for researchers and organizations.
Second step: structure the data into a graph. Use the NetworkX library (Python) to create nodes and edges. Each company is a node. Each transaction (invoice, contract) is an edge.
Third step: train a simple anomaly detection model. It doesn't need to be a GNN. An Isolation Forest algorithm can already identify companies with atypical behavior — such as issuing 500 invoices in one day.
Fourth step: share alerts with the police. The UN recommends that companies not attempt direct investigations but instead serve as bridges between data and authorities (UNODC, 2026).
Ethical and Technical Limitations
No technology is a silver bullet. AI against supply chain fraud faces three serious challenges.
First: algorithmic bias. If the model is trained only on data from past operations, it may learn to associate certain sectors or regions with suspicion. Interpol is developing an audit layer to mitigate this.
Second: privacy. Monitoring mass transactions raises legal questions. Safe Supply operates with judicial authorization and only analyzes public data or anonymized metadata.
Third: fraudster adaptation. When a pattern is discovered, fraudsters change their tactics. The model needs to be continuously retrained — the PF performs weekly updates.
Still, the numbers are promising. In 2025, the project identified 8 new routes. In 2026, it jumped to 12. The growth is exponential, because each mapped route feeds the model with new patterns.
Conclusion
Artificial intelligence will not end supply chain fraud on its own. But it is giving task forces an advantage they never had: the ability to see crime at scale. The 40% increase in fraud detection (Interpol, 2026) is not a cold number. It represents thousands of companies that are no longer anonymous figures in a global statistic.
The next step is to democratize this technology. Companies, auditors, and countries with fewer resources need access to ready-made models and shared data.
Related Articles
Related Articles
Hyperparameter Optimization with Hyperopt in 2026: Practical Guide
2026 practical tutorial: learn to optimize machine learning model hyperparameters using Hyperopt, with Bayesian search and result visualization.
Cyber Threat Detection with Graph Neural Networks in IoT Networks
How Graph Neural Networks detect attacks in IoT networks. Practical Python anomaly detection tutorial focusing on connected devices.
RL in the Supply Chain: Reinforcement Learning Tutorial for Optimizing Routes and Inventory in 2026
Practical tutorial on how to apply Reinforcement Learning (PPO) with Stable-Baselines3 and Gymnasium to optimize routes and inventory in supply chains, with...