Adversarially Robust Multigraph Neural Networks for Financial Crime Analysis

01 ABSTRACT

Fact: The paper proposes a structural attack and adversarial training framework based on node-edge incidence matrices, adapting projected gradient descent and projected randomized block coordinate descent, and introduces a cycle-based regularizer penalizing changes in global counts of directed 2- and 3-cycles. On synthetic pattern-detection and realistic AML datasets, learned attacks substantially outperform non-learnable perturbation baselines in reducing detection performance (PR-AUC), and adversarial training consistently improves robustness. Author's view: Multigraph neural networks are vulnerable to learned structural attacks, and adversarial training provides an effective defense.

02 KEY FINDINGS

1. Proposes the first scalable framework for structural attacks and adversarial training on multigraph neural networks, optimizing perturbations over node-edge incidence matrices.
2. Introduces a cycle-based regularizer penalizing changes in global counts of directed 2- and 3-cycles as a structural measure of attack plausibility.
3. Experiments show learned attacks significantly reduce PR-AUC, and adversarial training consistently improves model robustness.