Increasing complexity and digitalization of tax systems worldwide have changed the policy context in which tax administration occurs, making it possible to facilitate faster transactions, real-time returns, and data exchange between departments and other agencies. However, at the same time, such progress brought new opportunities for advanced tax fraud, including the use of shell companies, circular trading, under-invoicing, and organized cross-border evasion. Legacy systems for fraud detection – rule-based, point data mining solutions – have failed to adapt to these dynamic and complex threats. They frequently work in isolation and cannot scale across departments, nor capture the contextual information required to identify coordinated fraudulent behaviors across the network nodes of tax authorities.
In this paper, we propose a new multi-agent system architecture for coordinated fraud detection in the tax networks. MAS consists of autonomous, rational agents that can communicate, learn, and collaborate with other agents in a decentralized manner. Each agent of the proposed system is designed to serve a specific purpose, such as observing the behaviour of the taxpayer, detecting abnormal transactions, or exchanging alerts with other agents to work in concert for joint analysis. The framework is designed to operate in real time, incorporating new information in-flight and adjusting its detectors in response to learning, both within and between agents.
The proposed MAS-driven approach is organized following a layered architecture: (i) Detection Agents responsible of examining tax transactions and filings for inconsistencies or outliers using heuristic, statistical, and machine learning models; (ii) Coordination Agents, in charge of enabling the communication and consensus building of the detection agents through reasoning over the correspondent tax transactions affected, to obtain a complete view about the potential fraud across jurisdictions and tax categories; and, (iii) Learning Agents, aimed at keeping updating the detection models by learning from feedbacks over confirmed fraud cases, in order to be able to evolve in response to the emergence of new strategies for committing fraud. A novel feature of this architecture is the presence of a semantic ontology layer, which ensures the uniform structure and language of tax information, thereby enabling agents from various departments or agencies to communicate effectively with one another. It is crucial in a federated environment in which tax data is diverse and decentralized. Intra-agent negotiation protocols are also employed to resolve disputes and reach an agreement on whether a transaction or an entity is considered risky. The authors evaluated the proposed architecture through a simulation of a national tax network, utilizing both synthetic yet realistic data generated from anonymized tax filings and transaction logs, as well as known fraud scenarios. The MAS model was compared with conventional rule-based and centralized machine learning methods in terms of detection accuracy, false positive rate, scalability, and decision latency. The results show that the MAS-based system significantly enhances the system’s capability of uncovering various fraudulent patterns involving different tax types, regions, and filing channels. It further reduces response time through distributed processing and supports early alarm notification via proactive agent synchronization.
This paper is considered a contribution to an intelligent tax fraud detection framework, thanks to its scalability, adaptability, and decentralization, which meet the strategic objectives of modern revenue authorities. Its primary rationale is the integration of MAS into the current taxation system to improve efficiency, transparency, and cooperation across agencies. We hope that future work could also integrate blockchain for an audit trail, federated learning methods for privacy-preserving model updates, and human-in-the-loop for higher accountability.