Simulating and classifying behavior in adversarial environments based on action-state traces: an application to money laundering

Daniel Borrajo, Manuela Veloso, Sameena Shah

Introduction

Traditional Anti-Money Laundering (AML) systems primarily rely on transaction-based monitoring, which triggers alerts based on predefined, static patterns. However, money laundering is an inherently adversarial process where actors continuously adapt their strategies to bypass detection. This reactive approach often fails to capture the broader behavioral intent of sophisticated criminals. This paper proposes a shift toward behavioral trace analysis, focusing on sequences of actions and state changes to identify underlying goals, even under conditions of partial observability.

Relational Representation

To capture the complexity of financial networks, the researchers represent behavior as "traces"—ordered sequences of actions and the resulting environmental states. Unlike flat tabular data, this relational representation uses predicate logic to define connections between entities such as individuals, accounts, and corporations.

Table of predicates used to define the relational state space and their observability.

As shown in the predicate table, many critical states (e.g., money-laundering, criminal) are inherently unobservable to financial institutions. The challenge lies in inferring these hidden states from observable data like account-owner or transaction-origin.

Methods: The CABBOT Framework

The core of the proposed detection system is CABBOT (Classification of Agents’ Behavior Based on Observation Traces). CABBOT utilizes a relational k-Nearest Neighbor (kNN) approach to classify behavior. Rather than matching exact transaction amounts, the system employs distance metrics—most notably the Relational-based Distance (RIBL)—to compare the structural similarity of behavioral traces.

A key component of this framework is the ability to handle partial observability. In real-world scenarios, a bank only sees a subset of a customer's actions.

Table of actions within the simulation and their corresponding observability and domain type.

The action space includes both observable financial maneuvers (e.g., payroll, move-funds) and hidden adversarial actions (e.g., layering, perform-criminal-action). CABBOT is designed to classify the agent's goal based solely on the observable subset of these actions and the resulting state changes.

Synthetic Behavior Simulation

Due to the scarcity and sensitivity of real-world money laundering data, the authors developed a sophisticated simulator using Planning Domain Description Language (PDDL). Unlike random data generators, this simulator uses automated planning to give agents specific goals (e.g., "launder $10,000"). Agents then execute a sequence of actions to achieve these goals, creating realistic, interleaved traces of legitimate and illicit behavior. This allows for the generation of diverse training sets, including novel laundering typologies that the system has not previously encountered.

Key Results and Insights

The experimental results demonstrate the robustness of the relational approach:

  • High Classification Accuracy: CABBOT achieved near 100% accuracy in distinguishing "good" from "bad" behavior across various scenarios.
  • Rapid Convergence: Under full observability, the system often correctly classified a trace within a single observation step. Even with limited observability, CABBOT converged to the correct classification significantly faster than baseline methods like Decision Trees.
  • Generalization: The system successfully identified "unseen" tactics. For instance, a model trained on cash-based structuring could still identify digital money laundering because the underlying relational goals remained similar.
  • State-Difference Importance: The research found that "state-difference" metrics—measuring how the world changes after an action—often provided more predictive power than the action names themselves.

Conclusion

By moving from transaction-based alerts to relational behavioral traces, this work provides a path toward proactive AML detection. The combination of automated planning for data generation and relational learning for classification allows for an explainable system that can adapt to evolving adversarial tactics. Future research will focus on scaling these relational methods to handle the volume and noise of real-world global financial data.

Summary updated

Abstract

Many business applications involve adversarial relationships in which both sides adapt their strategies to optimize their opposing benefits. One of the key characteristics of these applications is the wide range of strategies that an adversary may choose as they adapt their strategy dynamically to sustain benefits and evade authorities. In this paper, we present a novel way of approaching these types of applications, in particular in the context of Anti-Money Laundering. We provide a mechanism through which diverse, realistic and new unobserved behavior may be generated to discover potential unobserved adversarial actions to enable organizations to preemptively mitigate these risks. In this regard, we make three main contributions. (a) Propose a novel behavior-based model as opposed to individual transactions-based models currently used by financial institutions. We introduce behavior traces as enriched relational representation to represent observed human behavior. (b) A modelling approach that observes these traces and is able to accurately infer the goals of actors by classifying the behavior into money laundering or standard behavior despite significant unobserved activity. And (c) a synthetic behavior simulator that can generate new previously unseen traces. The simulator incorporates a high level of flexibility in the behavioral parameters so that we can challenge the detection algorithm. Finally, we provide experimental results that show that the learning module (automated investigator) that has only partial observability can still successfully infer the type of behavior, and thus the simulated goals, followed by customers based on traces - a key aspiration for many applications today.
BibTeX
@inproceedings{borrajo2020simulating,
  author = {Daniel Borrajo and Manuela Veloso and Sameena Shah},
  title = {{Simulating and classifying behavior in adversarial environments based on action-state traces: an application to money laundering}},
  booktitle = {ICAIF},
  year = {2020},
  doi = {10.1145/3383455.3422536},
}
Figure