Paper Details

Name: Gurleen Kaur, Mandeep Kaur & Punam Rattan
Country: India

Year: 2025
Volume: Volume-12, Issue-2 (July-December)
Page Number: 272-287
DOI: https://doi.org/10.5281/zenodo.17798519

ABSTRACT
High-velocity transactions are made possible by mobile and portable wallets, but they also
increase the attack surface for low-signal, real-time fraud—often under stringent latency and
computation limitations. We present a low-compute, security-conscious hybrid learner for
wallet fraud detection that uses soft voting over a compact, domain-specific seven-feature
signature extracted from PaySim (step, amount, type, oldbalanceOrg, newbalanceOrig,
oldbalanceDest, and newbalanceDest) to couple logistic regression with a shallow decision
tree. In order to reduce false alarms and inference costs, the pipeline uses a rule-based
prefilter to exclude zero-information/system-generated records and explicitly handles extreme
class imbalance using transaction-type-aware SMOTE restricted to TRANSFER and
CASH_OUT. The trained artifact is serialized and sealed with SHA-256 to facilitate integrity
verification and governance checks, hence hardening deployment. With training and
inference performed on a low-resource laptop, the method achieves ROC-AUC 0.9917, F1
0.96 (precision 0.95, recall 0.98), and 0.96 accuracy on Pay Sim, indicating edge practicality.
While SHAP/LIME studies offer clear global and local explanations appropriate for
operations and compliance, benchmarks against LightGBM, Cat Boost, Random Forest, and a
CNN-LSTM demonstrate competitive or superior recall and ROC-AUC at significantly
reduced complexity. The contribution is a detection stack that is deployable, interpretable,
and governance-aligned while providing cutting-edge accuracy without the need for complex
models. We go over the drawbacks of evaluating synthetic data and provide strategies for
drift monitoring, live-stream validation, and privacy-preserving updates. These findings show
that on devices with limited resources, correctly designed, security-conscious hybrids may
offer dependable, real-time wallet fraud screening.
KEYWORDS:
High-velocity transactions, Logistic regression,Smote,SHA-256 , Cat Boost, Random Forest,
and CNN-LSTM