Securely Characterising Fraudulent Cycles in Financial Graphs

Abstract

With between 800 billion and 2 trillion US dollars laundered annually, which is up to 5% of the global GDP, money laundering undermines international financial systems. Crucially, these crimes are often intertwined with organised crime and violent illicit operations. Effectively identifying these illicit activities across transactional networks presents a critical dilemma: privacy regulations restrict data sharing, yet distributed data hinders fraud detection. Although graph-based strategies identify transaction cycles while respecting privacy regulations, existing methods cannot evaluate whether a cycle is benign or illicit before disclosing it. Many transactional cycles are benign, and revealing them raises privacy concerns. This paper addresses this gap by integrating the additive homomorphic Paillier cryptosystem into a pre-existing cycle detection propagation routine. Our protocol enables distributed institutions to sequentially aggregate per-node risk metrics without deciphering intermediate states. This allows nodes to evaluate collective cycle legitimacy homomorphically before revealing full cycles. We provide a theoretical evaluation of the framework’s privacy guarantees and space-time complexity under an honest-but-curious adversary model. Finally, we benchmark the implementation with a parallelised C++ implementation evaluated against a synthetic scale-free graph. While the protocol successfully achieves on-demand structural risk characterisation, the integration of 2048-bit Paillier operations introduces a substantial computational overhead of up to two magnitudes more than the baseline.