Financial crime, encompassing payment fraud, money laundering, identity theft, and insider trading, inflicts systemic damage on global economies. In the UK alone, fraud cost the banking industry $1.6 billion in 2024, and in April 2025 the UK government committed $161 million in quantum technology investment specifically to tackle crime, fraud and money laundering (World Economic Forum). Classical machine learning methods, whilst widely deployed, face structurally intractable challenges in this domain. Fraud datasets exhibit extreme class imbalance, with fraudulent transactions typically comprising less than 1% of records, non-stationary distributions, and high-dimensional feature spaces. These limitations reflect fundamental computational and statistical constraints that motivate exploration of different approaches.
Quantum Machine Learning (QML) offers a theoretically grounded alternative. Variational Quantum Circuits (VQCs) and quantum kernel methods embed classical data into exponentially large Hilbert spaces via parameterised quantum feature maps, enabling classifiers that operate in regions of function space inaccessible to polynomial-time classical algorithms. Quantum Generative Adversarial Networks (QGANs), and related quantum generative models such as Quantum Circuit Born Machines (QCBMs), utilise quantum superposition and entanglement to model complex, high-dimensional probability distributions, offering a principled mechanism for generating high-fidelity synthetic minority-class samples, directly addressing the class imbalance problem at its statistical root.
This PhD addresses three interconnected research challenges. First, it is not established which quantum feature maps and ansatz architectures are best suited to the distributional properties of financial crime data. Second, the question of whether QGAN-generated synthetic fraud data meaningfully improves quantum classifier performance, compared to standard quantum classifiers trained on imbalanced data alone, has not been rigorously answered on real hardware. Third, the adversarial threat surface of quantum financial crime detection systems remains uncharacterised: it is unknown whether VQCs and quantum kernel classifiers are inherently more or less robust than classical equivalents to adversarial perturbations crafted specifically in the financial crime domain.
The programme of work intends to proceed in three phases: (a) the design and execution of hardware-aware VQC and quantum kernel classifiers on IBM Quantum hardware, systematically investigating the role of data encoding strategies (amplitude, angle, and IQP-style encoding) and entanglement topology on classification performance for financial fraud data, (b) QGAN architectures will be trained and executed on quantum hardware, targeting the generation of statistically faithful synthetic fraudulent transaction data, and (c) a formal adversarial threat model will be developed for the quantum detection pipeline, covering input-space perturbation attacks, parameter-shift gradient leakage, and model extraction, with quantum-noise-aware defences evaluated on hardware.
The expected contributions include: (i) a hardware-validated, domain-specific framework for quantum feature map and ansatz co-design for financial crime data, (ii) the first systematic evaluation of QGAN/QCBM-augmented versus standard
quantum classifiers for fraud detection on real quantum processors, (iii) a formal adversarial threat taxonomy and mitigation framework specific to quantum financial crime detection systems, and (iv) a cross-platform benchmarking spanning superconducting and trapped-ion architectures that provide empirically grounded guidance on hardware selection for financial QML applications.
This interdisciplinary project sits at the intersection of quantum computing, machine learning, cybersecurity, and financial technology, and offers the opportunity to contribute to an emerging field with significant scientific and societal impact.
