Forklarbar kunstig intelligens for å bekjempe hvitvasking

Offentlig prosjektsammendrag

Numerous banks have in recent years faced substantial fines due to ineffective anti-money-laundering (AML) controls. One such control is based on transactions monitoring systems (TM systems), designed to detect suspicious money transfers. Today's TM systems are predominantly based on defining a set of if-then rules with predefined thresholds. If a threshold is exceeded, the TM systems trigger alerts that are forwarded to AML specialists for evaluation. Unfortunately, with an estimated false-positive rate ranging between 95 to 98%, banks find themselves struggling to comply with financial regulation. Although deep learning has been proposed as a remedy, a major obstacle must be overcome, namely the lack of interpretability within deep learning models. Financial institutions are obligated to be able to explain the reasoning behind a decision to their clients, regulators, and other stakeholders. Deep learning models, however, typically involve thousands if not millions of parameters that do nonlinear computations, making it difficult for humans to follow the reasoning behind a model's predictions. Deploying a deep learning model into a live environment without a comprehensive understanding of its decision-making process places banks at risk of discriminating against their customers. The consequences of such biases include legal liability, reputational damage, and loss of customers' trust. Recent advances in interpretable deep learning theory offer some promising tools in this regard. Techniques such as saliency maps and gradient-based methods offer avenues for extracting meaningful interpretations. Yet another interesting approach is the concept of 'attention as explanation'. In this project, our aim is to build effective and interpretable deep learning algorithms to combat money laundering. We will do so by combining theory on interpretable deep learning with theory on financial crime.