Multi-Adapter Qlora for Cross-Domain Financial Fraud Detection

Abstract

Most fraud detection systems are developed for a single domain at a time: a bank uses one model for credit card fraud, an insurer uses another for claims, and each requires its own large labelled dataset while offering no explanation for its decisions. This paper tests whether a single lightweight language model can instead serve several fraud domains at once. We attach three small, independently trained QLoRA adapters, one each for credit card fraud, insurance claims fraud, and antimoney laundering (AML), to one frozen, quantised TinyLlama1.1B backbone. During inference, the appropriate adapter is loaded onto the backbone, and the model returns a fraud verdict, a plain-language rationale, and a calibrated risk score, all from a single GPU. The approach performs competitively on credit card and AML detection using far less training data than standard classifiers, but fails outright on insurance fraud, collapsing into classifying every case as fraudulent. We trace this failure to a specific cause: insurance fraud depends on information spread across a customer's claim history, which cannot be captured when the model sees only one record at a time, so this negative result is reported as a concrete boundary case for the approach. We further calibrate the model's risk scores, measure adapter hot-swap latency, test blending two adapters together, and study how adapter size and training data volume affect performance. The contribution is a practical, reproducible evaluation of where multi-adapter QLoRA works and where it needs further architectural support, particularly for domains where fraud signals are relational rather than contained in a single record.