A FEDERATED EDGE SMALL LANGUAGE MODEL FOR ADAPTIVE FINANCIAL COUNSELING AND FRAUD-AWARE SPENDING RECOMMENDATION WITH REGULATORY-GRADE EXPLAINABILITY

Abstract

Modern mobile banking applications provide a record of how customers earn, spend and save money, whereas digital banks continue to provide universal advice on budget planning and alert on frauds by rules. Cloud large language models can give fluent financial guidance and recommendations, but their application in such an environment is prohibited by the data minimization principles of the General Data Protection Regulation, lack of compliance with the transparency and suitability requirements of the MiFID II, and latency issues. In this paper, we describe the implementation of a federated edge small language model (SLM) that is run within the mobile banking application and provides personalized spending guidance and budget and fraud alerts without transferring transaction histories anywhere else than the mobile device. Each client trains a low-rank adapter of an instruction-tuned compact SLM model using their local transaction histories, and only adapter parameters are uploaded using secure aggregation with clipping, Gaussian noise, and dropout-resilient masking. A meta-learned initialization shared among clients solves the cold-start problem and allows getting a workable advisor after several local transactions. The recommendation issued by the federated model has two explanation channels: Shapley attributions of interpretable spending features, and the smallest possible change in behaviour leading to the desired financial goal. The jurisdiction-aware aggregation function punishes disagreements of client models in order to keep the federated advisor aligned with the financial regulatory framework applicable in each particular country. We evaluate the federated edge model on a federation of 3.2 million labelled transactions coming from six consenting organizations, and Bank Account Fraud and IEEE-CIS fraud benchmark datasets. Federated edge model is able to detect fraud with 98.7% accuracy, 0.911 F1-score and 0.984 AUC ROC score while preserving the privacy budget at \epsilon=2.0; it gives stable, sparse and human-friendly explanations accepted in 87.4% of audited cases. Median latency on a mid-range mobile phone is only 43 ms, so the advisory engine works completely offline.