Hybrid quantum network flags suspicious accounts with fewer false alarms

More than 95% of alerts generated by current anti-money laundering and trade surveillance systems prove false, demanding exhaustive manual review by human analysts. To address this inefficiency, researchers from Vanguard, The Washington Institute for STEM, Entrepreneurship and Research (WISER), and Capgemini benchmarked four quantum and hybrid quantum-classical classifiers against classical models in a synthetic pump-and-dump detection task.

The work demonstrates a hybrid quantum neural network, executed on the ibm_fez superconducting processor, achieved a suspicious-class F1 score of 55.4-59.8%, performing on par with well-tuned classical models while using only five qubits. The team reports this suggests “considerable potential” for quantum approaches to improve trade-surveillance analytics.

Quantum-Hybrid Models Benchmark for Pump-and-Dump Detection

The hybrid quantum neural network achieved an F1 score of 55.4-59.8% when executed on the ibm_fez superconducting processor, exceeding the performance of six out of eight classical models tested. This result indicates the potential for improved accuracy in identifying pump-and-dump schemes, a critical need given that current anti-money laundering systems generate alerts where over 95% prove to be false positives, requiring extensive manual review. The HQNN’s ability to achieve this level of performance using only five qubits represents a step toward practical quantum applications in finance.

Compared to its larger 20-qubit quantum neural network counterpart, which achieved a 42.0% F1 score, the hybrid model reduced false alarms by over 50% while maintaining a similar detection rate, demonstrating a substantial efficiency gain. The research team found that, at a 15% false-positive rate, the HQNN correctly identified one in every two suspicious accounts, a rate comparable to well-tuned classical models like XGBoost, which reached 61.4% F1.

While the XGBoost model still held a slight edge on this benchmark, the performance of the five-qubit hybrid circuit suggests that further development could close the gap. The team also evaluated a fidelity-based quantum k-nearest-neighbours classifier (QKNN) and a quantum support vector machine (QSVM), which achieved F1 scores of 43.5% and 41% respectively, falling short of the classical XGBoost benchmark.

XGBoost Outperforms Quantum Classifiers on Synthetic Data

XGBoost achieved a 61.4% F1 score. The QKNN reaches 43.5% F1 (44.4% under noise), above its classical counterpart (35.6%). The researchers noted that while the HQNN showed competitive results, no quantum or hybrid model exceeded the capabilities of the best classical model on this benchmark, consisting of 800 accounts for training and 200 for testing.

HQNN Reduces False Alarms Compared to 20-Qubit QNN

The hybrid quantum neural network (HQNN) generated fewer false alarms than a comparable 20-qubit quantum neural network, flagging genuinely suspicious accounts in simulated and real-world tests. Achieving an F1 score of 55.8% on the ibm_fez processor, the HQNN outperformed six of eight classical configurations, while using only five qubits and requiring approximately a quarter of the training time of its larger counterpart. This reduction in false positives translates to a significant decrease in the workload for human analysts currently tasked with manually reviewing over 95% of alerts generated by existing systems.

On ANOVA feature sets, the HQNN’s best F1 score reached 57.9%, surpassing logistic regression (41.3%) and a three-layer neural network (52.4%), though remaining behind the performance of a one-layer neural network (59.8%) and XGBoost (61.4%). The HQNN’s ability to detect more suspicious accounts while reducing benign flags suggests a more efficient use of resources in trade surveillance. The study reports, highlighting a potential advantage in computational efficiency.

Five-Qubit HQNN Performance on IBM_fez Hardware

The hybrid quantum neural network (HQNN) maintained detection performance on actual quantum hardware, achieving F1 scores of 55.8% on the ibm_fez processor using ANOVA feature selection, a result comparable to the strongest classical models tested in the study. The HQNN’s ability to function effectively on limited quantum resources suggests a pathway toward practical applications of quantum machine learning in financial analytics.

While the comparison does not definitively prove the five-qubit circuit adds unique value beyond a similarly sized classical layer, the HQNN reached a comparable performance level to the best classical models, using thresholds fixed on training data alone, both in simulation and on real hardware. This suggests the representation and training procedure, rather than noise, are the primary limiting factors at these circuit depths.

A key advantage of the HQNN lies in its practicality; its shallow, five-qubit circuit is easier to run on current hardware, a benefit that will likely increase as the gap between available and required qubit counts widens. The team acknowledges that further work is needed to confirm this advantage by replacing the quantum circuit with a classical layer of equivalent size and retraining the model. But the initial results demonstrate a promising direction for quantum-enhanced trade surveillance.

QKNN Achieves Competitive F1 Score with Noise Simulation

The clustering-based quantum k-nearest-neighbours classifier (QKNN) demonstrated resilience even when subjected to simulated noise, a key factor for real-world deployment. Evaluations using the FakeFez noise model revealed an increase in the QKNN’s recall across both feature representations. This contrasts with purely classical k-nearest-neighbours, used as a control, and indicates a potential advantage in maintaining performance under imperfect conditions.

The HQNN is the strongest quantum-hybrid model, reaching 54.5-57.9% F1 on an ideal simulator and 55.4-59.8% on ibm_fez, ahead of six of the eight classical configurations. The classical network handles the input dimension, allowing for this reduction in quantum resources.

This hardware execution, alongside simulated noise testing of the QKNN and QSVM, represents a step toward validating performance on actual quantum hardware. However, the researchers note several caveats, including the use of a single training run for quantum models versus a ten-seed sensitivity analysis for classical models, and the synthetic nature of the benchmark dataset with its 20% suspicious rate, a figure far exceeding the less than 1% rate typical of production anti-money laundering systems. “Every method exposed to noise tolerated it,” the study reports, indicating a degree of robustness across the tested algorithms.

QSVM Performance Lags Behind Other Quantum Methods

The quantum support vector machine (QSVM) demonstrated performance trailing other quantum approaches in a benchmark designed to detect pump-and-dump financial schemes. This difference highlights the varying efficacy of different quantum algorithms when applied to the same financial fraud detection task. The QSVM does not exceed 41%.

Unlike the HQNN, which adapts through training of its circuit parameters, the QSVM’s kernel remained static, limiting its capacity to refine its detection criteria. This benchmark, using 20 features per account and 40 suspicious accounts in the test set, revealed that even the best quantum models did not surpass classical methods like XGBoost, which achieved a 61.4% F1 score.

The QKNN, another quantum classifier, also underperformed, trailing the best classical models by around 17 points of F1. These results suggest that, on this synthetic benchmark, quantum algorithms have yet to demonstrate a clear advantage over established classical techniques.

Financial Crime Drives Need for Advanced Surveillance

The global financial system loses an estimated $800 billion to $2 trillion annually to money laundering, a figure representing two to five percent of global GDP. This scale of illicit finance drives a constant need for improved detection methods, particularly as traditional rule-based systems struggle with the volume and complexity of modern transactions. Standard classifiers often misinterpret fraudulent activity as noise due to the extreme class imbalance inherent in financial crime data, where illicit transactions represent less than 0.1% of all activity.

Quantum Support Vector Machines (QSVMs) offer a potential solution by using quantum feature maps to embed classical data into a complex quantum Hilbert space. This approach aims to evaluate non-linear similarities that are classically intractable, providing an alternative method for identifying suspicious patterns. The construction used in this work encodes one feature per qubit, meaning the number of qubits equals the number of features being analyzed.

Classical Machine Learning Limitations in AML Systems

Financial crime currently costs the global economy between $800 billion and $2 trillion each year, a figure driven in part by the limitations of existing detection systems. This challenge is compounded by the complex, high-dimensional nature of financial data and the relational links within it.

Recent approaches to address these shortcomings include quantum k-nearest neighbours (QKNN), an instance-based classifier that predicts class based on similarity to training data. This framework maps classical samples into quantum states and uses quantum fidelity to construct a similarity structure, preserving local geometric relationships through a k-nearest-neighbour graph and a deterministic merging strategy.

High False Positives Plague Current AML Rule-Based Systems

This high rate of false positives creates a significant bottleneck for financial institutions, diverting resources from genuine investigations into illicit financial activity. The sheer volume of alerts necessitates a re-evaluation of current detection methods, as existing rule-based systems struggle to adapt to evolving criminal tactics. This represents a potential order-of-magnitude reduction in the burden on human analysts, assuming similar performance holds in real-world scenarios. The QKNN also showed promise, improving upon a classical k-nearest-neighbours control and maintaining performance even with simulated device noise.

The practical implications of these findings extend beyond purely algorithmic improvements, as a reduction in false positives directly translates to cost savings and increased efficiency for compliance teams. “AML and trade-surveillance review is still largely manual, with more than nineteen in twenty rule-based alerts turning out to be false,” the paper notes, highlighting the scale of the current challenge. The ability to filter out legitimate transactions more effectively allows investigators to focus on higher-risk cases, ultimately strengthening defenses against financial crime.

More like this

Quantum AlgorithmsResearchers Map How Dipoles Tune Polaron Binding Energy

Quantum Research NewsFour photon ‘shapes’ boost quantum network capacity

Quantum AlgorithmsGeorgia Institute of Technology Finds Limits to Ground State Preparation

Quantum AlgorithmsResearchers Propose New Path to Large Quantum Circuits