On the Applications of Machine Learning for Alleviating Challenges in the Financial Crime Domain

Abstract

The rapid digital transformation within the financial sector is driven by the integration of technologies for increasing operational efficiency and achieving sustainable development. The development is further accelerated through the leverage of vast volumes of data, reflected by the unprecedented innovation in several financial domains. However, the digitalisation of products and services are maliciously exploited by criminal actors engaging in financial crime. In particular, there has been an incline in card fraud and money laundering offences, which are yet expected to persist as challenges for the authorities and financial institutions in the foreseeable future. Due to the drastic consequences of financial crime on society and the economic system, it is imperative that efforts for combatting financial crime need to be strengthened. The current landscape of card fraud detection and anti-money laundering (AML) is characterised by the dependency on rule-based systems. These systems are constructed on pre-defined criteria derived from domain expertise and regulatory guidance. However, the simplicity of these rules often leads to a high volume of alerts that necessitate further investigation, a process that is both time-consuming and resource-intensive. Moreover, criminals continuously devise new strategies for committing card fraud and money laundering, rendering the reliance on such rule-based systems infeasible. More recently, there has been a growing interest in employing machine learning algorithms for identifying suspicious transactions and overcoming limitations in traditional financial crime detection models. Despite a growing body of academic literature exploring machine learning techniques for fraud detection and AML, it is not obvious how the research body contributes to the integration of such systems by financial institutions. Additionally, the works in scholarly articles frequently rely on small or simulated datasets, which are not representative the complex data environments managed by financial organisations. To this end, the main objective of this thesis is to address challenges present in the development of advanced fraud and money laundering detection frameworks, with a particular focus on a number of aspects. The relative scarcity of illicit financial transactions in contrast to legitimate transactions, in addition to absence of the reliable ground of truth in the labelling of data instances give rise to the representation of data in alternative setting. Therefore, in this research a semi-supervised learning framework for classifying entities is introduced, designed to help in identifying customers suspected of money laundering based on their historic attributes, while reducing the impact of class imbalance and poor data labelling. To address the disparity between proprietary data of financial institutions and unrepresentative data typically used in academic research - typically due to confidentiality concerns — a tabular generative adversarial network (GAN) was designed to generate realistic synthetic tabular data that emulate the statistical properties of real data. The robustness of generative models against privacy attacks was further analysed, indicating that sophisticated reconstruction attacks on synthetic datapoints pose a threat in unveiling sensitive information. Moreover, this thesis extends the discussion to practical concerns that are frequently neglected in scholarly research, such as data drift and the interpretability of machine learning models. The findings of this research suggest that machine learning holds promise in alleviating the impacts of financial crime. Nevertheless, obstacles for adopting artificial intelligence and machine learning models by financial institutions need to be progressively addressed.