FedGuard: An Explainable Centralized Fraud-Detection Baseline on PaySim and IEEE-CIS
Fraud detection is difficult because fraudulent transactions are rare, data differ across domains, and strong predictive models can be hard to review. We present FedGuard, an explainable centralized baseline evaluated separately on two public benchmark domains: PaySim, a synthetic mobile-money dataset with 6.36 million transactions and 0.13% fraud, and IEEE-CIS, an anonymized card-not-present e-commerce dataset with 590,540 labeled transactions and 3.5% fraud. The records were split into train, validation, and test partitions before any preprocessing was fitted. Train-only preprocessing and sampling were used in four model families and 195 candidate fits. Thresholds were selected using validation data to maximize recall subject to an FPR no greater than 1%. On the held-out test split, the selected LightGBM model reached AP 0.997968, recall 99.76%, and FPR 0.003567% on PaySim; for IEEE-CIS, AP was 0.756757, recall 68.80%, and FPR 1.017758%, slightly above the validation target. TreeSHAP contributions for the selected tree model sum with the baseline value to the model raw output; they are not calibrated probabilities. The codebase includes a 55-case software test suite, which checks implementation behavior rather than trained-model performance. This work does not implement federated learning, differential privacy, or secure aggregation.
Authors
- Mohit Tiwari (ORCID: https://orcid.org/0000-0003-1836-3451)
- Anvi Tyagi (ORCID: https://orcid.org/0009-0007-4091-9897)
- Riyanshu (ORCID: https://orcid.org/0009-0007-6984-2782)
- Naman Dugar (ORCID: https://orcid.org/0009-0005-2119-4128)
Institutions
- Bharati Vidyapeeth's College of Engineering, Delhi (IN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-19
- DOI
- https://doi.org/10.21203/rs.3.rs-11739412/v1
- Primary Topic
- Imbalanced Data Classification Techniques
- Type
- preprint