Digital payments have grown at a remarkable pace,
and fraud has grown with them. Rule-based detectors served
banks well for years, but they struggle to keep up with attackers
who keep changing their tactics. Machine learning has helped,
yet many deployed models still raise too many false alarms, adapt
slowly to new behaviour, and do not carry over well from one
transaction setting to another. This paper presents FinShield AI,
a hybrid framework that pairs a Bidirectional Long Short-Term
Memory (BiLSTM) network with a Graph Attention Network
(GAT) and places an Isolation Forest in front of both as a
lightweight anomaly pre-filter. The BiLSTM reads how each
customer’s transactions unfold over time, while the GAT looks
at how users, merchants, devices and IP addresses are linked
to one another, so the model sees both the temporal and the
relational side of fraud. On the IEEE-CIS Fraud Detection and
PaySim datasets, FinShield AI reached 98.74% accuracy, 97.83%
precision, 97.41% recall and a 97.62% F1-score, ahead of every
baseline we tested.

Fraud Detection, Digital Payments, BiLSTM, Graph Attention Networks, Isolation Forest, Anomaly Detection, Real-Time Classification, Deep Learning