Federated and Drift-Resilient Machine Learning for Financial Fraud and Distress Detection

Abstract

Federated learning allows multiple financial institutions to train a shared fraud-detection or financial-distress model collaboratively without any institution exposing its raw transaction or account data to the others, a privacy-preserving property that has made it an attractive architecture for cross-institutional financial machine learning. That architecture, however, was designed around an assumption that becomes increasingly fragile the longer a federation operates in production: that each participating institution’s local data distribution remains reasonably stable, so that averaging every institution’s model updates together each communication round continues to produce a sensible shared global model. In practice, individual institutions experience concept drift on different timelines and for different reasons a regional economic shock affecting one institution’s borrower population, a coordinated fraud campaign targeting one institution’s card portfolio specifically, a new product line at one institution shifting its transaction mix so a federation’s participating institutions rarely drift in unison, and a global aggregation rule that treats every institution’s update identically regardless of how reliable that update currently is risks corrupting a well-functioning shared model with one institution’s idiosyncratic, currently unreliable local signal. This paper develops Federated Learning with Adaptive Resilience to drift Events (FLARE), a framework that adds two mechanisms to a standard federated learning pipeline without weakening its core privacy guarantee: each participating institution runs local concept-drift detection on its own data and shares only a drift status and severity signal, never raw data or gradients derived from suspect data, with the central aggregation server; the server then down-weights or temporarily excludes updates from institutions currently reporting severe, unconfirmed drift when computing the shared global model, protecting the federation’s other members from a single institution’s idiosyncratic instability; and an institution confirmed to be experiencing genuine, sustained local drift is offered an optional personalized local fine-tuning branch layered on top of the shared global model, so that institution can adapt to its own changed conditions without waiting for federation-wide consensus and without dragging the shared global model off course for everyone else. Following a design-science research approach, the framework is illustrated, not validated, through two comparative tables, a hypothetical case scenario, and four illustrative figures, with every quantitative value explicitly and repeatedly labeled as an illustrative estimate constructed from the cited literature’s directional findings rather than as measured data from a deployed federation. The paper concludes with a discussion of FLARE’s practical implications for financial institutions participating in cross-institutional federated learning, its assumptions and limitations, and directions for future empirical research.