scission: A Hybrid Classical-AI and Quantum Pipeline for Fraud-Ring Detection in Transaction Graphs: Design, Implementation, and Empirical Evaluation
Muhammad Asghar Ali
Coordinated financial fraud manifests as small, dense communities embedded in large transaction graphs. Isolating such a community is a combinatorial graph-partitioning problem of the kind often proposed for quantum optimisation, yet near-term quantum devices hold only tens of qubits while real graphs hold millions of accounts. We present scission, a hybrid pipeline built and evaluated under a design-science methodology, whose organising principle is that classical AI compresses the problem so a quantum solver can cut its combinatorial core. Classical anomaly scoring and a greedy reduction shrink the graph to a qubit-budget subgraph; a two-community modularity QUBO then isolates the ring, solved interchangeably by exact enumeration, simulated annealing, or a dependency-free statevector QAOA. We define an experimental protocol with explicit research questions, ground-truth metrics, and reproducibility controls. On a synthetic planted-ring benchmark (44 accounts reduced to 14 qubits) every backend attains the exact optimum, recovering all six ring members with one false positive (F1 = 0.923). We report no quantum advantage at feasible scale, formalise the pipeline's binding constraint as a reduction-recall bound, and give a structured research agenda toward real-data, real-hardware, and adversarial validation. We argue the contribution is architectural: a neutral QUBO interface behind which the solver improves as hardware matures, and an honest, reproducible account of where value does and does not yet arise.
•Aug 10, 2026•0 citations