Developing Agnostic Network AI Models for Financial Crime Detection

Financial crime

Financial crime encompasses illegal activities involving financial processes for personal or organizational gain. Such activities generate illicit funds that require laundering to be integrated into the financial system. For example, criminal networks engaged in human trafficking often rely on the financial sector to fund recruitment and transportation, bribe public and private actors, and conceal profits derived from exploitation. This reflects a significant and rapidly growing trend: a global rise in non-violent crimes, particularly financial and cyber-dependent offenses.

AI Functionality and the Role of Data

Usually hidden and complex, data on financial crime are limited, imbalanced, or restricted. Still, AI models can perform key tasks such as prediction, detection, and generation. To do so, they must be trained: their internal parameters are adjusted based on observed data until they can reliably perform a specific task.

The effectiveness of an AI model depends on both the quality of the data and the nature of the task. These factors determine how complex the model needs to be. For instance, a simple yes/no classification task (e.g., detecting whether an image contains a dog) requires far less complexity than generating images or answering open-ended analytical questions.

Understanding this balance is crucial when applying AI to financial crime detection—where data are imperfect and decisions carry real-world consequences.

Machine learning lies at the core of modern AI. Broadly, it can be understood in two ways: rule-based and statistical. In rule-based approaches, systems follow predefined instructions—for example, “If a transaction exceeds $10,000, flag it”. While straightforward, such systems quickly become rigid, difficult to scale, and unable to adapt to new situations without manually adding more rules.

To overcome these limitations, statistical machine learning was introduced. Instead of relying on fixed rules, models learn to recognise patterns from data. For instance, if a model has been trained on many examples of dogs—including those in unusual settings like a snowy mountain or a crowded city street—it can still correctly identify a dog in a completely new background. This is because its predictions are shaped by patterns learned across many diverse observations, rather than a fixed set of instructions.

Today, most AI systems rely on this statistical approach. In practice, when we refer to “machine learning,” we imply statistical machine learning.

When is statistical machine learning the right tool? In practice, three conditions should be met: (1) there is an underlying pattern, (2) the pattern is not fully deterministic or easily defined by explicit rules, and (3) sufficient data exists to learn from.

If a problem is fully deterministic—such as calculating the sum of two numbers—using AI is unnecessary and inefficient. Similarly, if experts can clearly define patterns (for example, known financial crime indicators), simpler methods like rule-based systems or pattern matching may be more appropriate than machine learning.

Machine learning becomes valuable when patterns are complex, uncertain, or difficult to formalize—yet still observable in data.

In this context, “learning” refers to the process of reaching a stable state where the model’s parameters capture these patterns. “Training” is the iterative process of adjusting those parameters using data until the model performs reliably.

AI in Financial Crime Detection

To apply AI to financial crime detection, we must first define how we model the data—this determines the model’s scope, complexity, and task (e.g., detection or prediction). One powerful approach is to represent financial transactions as a network: entities (such as individuals or accounts) are nodes, and transactions between them are edges.

This perspective allows us to capture higher-order relationships. Financial crimes rarely occur in isolation; they often emerge within hidden and complex subnetworks. Focusing on individual transactions alone is therefore insufficient.

However, this understanding of network-structured financial crimes presents key challenges: the data are highly imbalanced (a few suspicious cases among millions of normal ones), reliable ground truth is limited, and the data themselves are sensitive and not publicly available.

You might wonder whether general-purpose AI models—such as ChatGPT or other large language models—can be applied here. In practice, this is difficult. These models are not designed to handle network-structured data, and more importantly, strict privacy and regulatory constraints prevent financial institutions from sharing sensitive transaction data with external systems.

Financial Crime Detection in the COMCRIM AI Work Package

Within the COMCRIM AI PhD project, the goal is to develop AI models for detecting financial crimes while addressing key real-world constraints, including data imbalance, limited ground truth, restricted access, and high sensitivity.

To this end, we formulate financial crime detection as an anomaly detection problem, focusing on identifying rare and unusual patterns in transaction networks. This approach enables the capture of both known and previously unseen behaviours, as well as higher-order structural patterns that are not observable when analysing isolated transactions or sequential data alone.

For example, an account exhibiting rapid inflow and outflow of funds reflects a well-known money laundering pattern. Similarly, a business account with unusually high incoming funds relative to its interactions with other accounts (or subnetworks) may indicate irregularities, even if the exact cause remains unclear.

Anomaly detection is designed to flag such deviations from normal behaviour, regardless of whether the pattern is already known. This makes it particularly valuable in financial crime settings, where new tactics continuously emerge. Moreover, anomaly detection methods share common principles across domains—whether applied to networks, images, or video—making them both flexible and widely applicable.

To address limited data availability, we leverage pre-trained models. Rather than training from scratch, we fine-tune existing models, significantly reducing data requirements. In parallel, we develop anomaly detection methods tailored to network-structured data, enabling more flexible and robust learning under severe class imbalance (e.g., millions of normal transactions versus a small number of unusual ones).

A key challenge, however, is explainability. Detecting an anomaly is not sufficient; banks, investigators, and other stakeholders need to understand why a case is flagged. This requires AI models to go beyond assigning abstract anomaly scores and instead provide clear, human-interpretable explanations of suspicious patterns, thereby facilitating the verification or falsification of evidence.

Within the COMCRIM AI track, we develop domain-agnostic network AI models that operate across diverse networks, including financial transactions, human trafficking, and other illicit structures. These models are not constrained by predefined patterns, enabling the detection of potential financial crime signals in a broad sense. 

Our goal is to provide reliable anomaly detection alongside interpretable explanations, supporting the identification and understanding of suspicious patterns.

You will be hearing more from us in the near future.

*Yassin Mohamadi is a researcher in Generative AI at the University of Amsterdam and part of the COMCRIM project.