Archer has completed the latest stage of its quantum machine learning fraud detection project, including tests on both a quantum simulator and a commercial quantum computer.
Its quantum neural network model was benchmarked on a public financial fraud dataset with more than 280,000 transaction records. The model was designed to work within current quantum computing limits, using dimensionality reduction and data balancing techniques before testing.
In the simulator environment, the early model matched the strongest classical models in Archer's benchmarking. It correctly identified 118 fraudulent transactions, missed 30, and produced one false positive.
False positives are a key issue in fraud detection because they can trigger unnecessary reviews and disrupt customers. Archer highlighted the low false-positive rate in the simulator as one of the more notable results from this phase of the project.
The model was also tested under simulated quantum noise. The system remained stable at low noise levels and showed only minor degradation at moderate levels, while performance fell more sharply at higher noise levels.
This gives Archer a clearer view of the hardware conditions needed for practical use of quantum machine learning in fraud screening. It also reduces some of the technical uncertainty around how such models might perform outside idealised test conditions.
Separate validation took place on IQM Garnet, a 20-qubit superconducting quantum computer accessed through AWS Braket. In that hardware test, the model detected 18 of 19 fraudulent transactions in the test set.
Archer said the real-hardware run produced a higher false-positive rate than the simulator tests. Even so, it described the result as evidence that the model can operate on current commercial quantum hardware as well as in simulation.
Benchmarking work
Archer used a staged experimental framework to determine the model configuration. The process included qubit-selection studies, feature-map optimisation, comparisons with classical machine learning methods, and analysis of the effect of quantum noise.
According to Archer, this work established a repeatable benchmarking framework and identified a quantum architecture that performed well on the chosen fraud dataset. It also said the project has not yet shown a clear edge over leading classical artificial intelligence methods.
Archer's progress comes as banks and quantum computing groups continue to examine fraud detection as one of the more practical commercial uses for quantum machine learning. It pointed to work by Quantinuum with HSBC and a separate collaboration between Intesa Sanpaolo and IBM as examples of similar efforts elsewhere in the market.
These programmes reflect broader interest from financial institutions in tools that can analyse large transaction volumes while limiting missed fraud and unnecessary alerts. For quantum computing companies, fraud detection offers a defined test case with measurable outcomes against existing machine learning systems.
Research phase
The latest milestone follows Archer's earlier completion of dataset preparation before the project moved into simulation and benchmarking. This phase was carried out on a prepared research dataset and with a selected set of comparison models, rather than under production banking conditions.
More testing is required before any commercial path can be assessed. Archer said this would include larger datasets, additional classical benchmarks, repeated trials, and further hardware validation.
The work forms part of Archer's broader activity in quantum computing, sensing, and medical diagnostics. The company operates in the semiconductor sector and has been developing chips related to quantum applications.
Dr Simon Ruffell, chief executive officer of Archer, said the results marked a useful step in testing whether quantum machine learning can be applied to fraud detection under current technical limits.
"These results have demonstrated that QML approaches can deliver strong fraud detection performance while operating within the constraints of current quantum computing systems. The simulator results were solid, particularly the very low false-positive rate, and the successful execution on real quantum hardware is an important validation step. The research collaboration agreement with the CSIRO forms part of Archer's strategy to investigate practical applications of quantum computing technologies and support future commercialisation opportunities in data-intensive industries. Fraud detection is a relevant use case for QML because banks and payment providers must analyse large volumes of transaction data quickly, while reducing both missed fraud and false alerts," Ruffell said.