Federated Credit Card Fraud Detection Under Simulated Operating Conditions: Encrypted Aggregation, Partial Participation and Byzantine Resilience
International Journal of Advanced Computer Science and Applications · 2026 · 17 (10)
DOI: 10.14569/ijacsa.2026.0171092 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Advanced Computer Science and Applications, 2026. Verfasst von El Mehdi Adnani und 1 weiteren Autor:innen.
Abstract (Originalsprache)
Credit card fraud is rare, costly and continuously evolving, and the data required to detect it are distributed across institutions that cannot share them. This study presents a privacy-preserving federated framework in which several banks jointly train a fraud detector without exchanging raw transactions, and it quantifies the cost of each privacy and robustness mechanism. The evaluation uses ten simulated institutions obtained by partitioning a public benchmark of 284807 transactions. A federated autoencoder learns legitimate spending without fraud labels, and a nearest neighbour classifier operating on the per-feature reconstruction errors provides the decision, using labelled error vectors. The placement of this classifier determines the trade-off between privacy and accuracy. Fitted on pooled error vectors, it attains an F1-score of 0.8625 over five seeds, comparable to the centralized model at 0.8667. Fitted locally at each bank, it attains 0.7254 with no exchange of data, and a head trained by federated averaging, which exchanges only parameter updates, attains 0.7655, close to a federated supervised network on raw features at 0.783. On the same split, label-free detectors remain below 0.45 and a centralized random forest reaches 0.835. CKKS homomorphic encryption conceals every update from the aggregation server and reproduces plaintext averaging to within 1.8e-7, at 25.6 times the ciphertext volume and 1.011 times the computation. Robust aggregation operates on plaintext updates and is evaluated separately: the coordinate-wise trimmed mean sustains an F1-score of 0.8402 when a fifth of the institutions mount a gradient inversion attack, and fails at two fifths. A participation rate of 60 percent remains within seed variation of full participation while reducing traffic by 40 percent. These results provide a consortium with a quantified trade-off between privacy, robustness and detection quality.
Zitieren
APA
Adnani, E. M., & Erritali, M. (2026). Federated Credit Card Fraud Detection Under Simulated Operating Conditions: Encrypted Aggregation, Partial Participation and Byzantine Resilience. International Journal of Advanced Computer Science and Applications, 17(10). https://doi.org/10.14569/ijacsa.2026.0171092
Vancouver
Adnani EM, Erritali M. Federated Credit Card Fraud Detection Under Simulated Operating Conditions: Encrypted Aggregation, Partial Participation and Byzantine Resilience. International Journal of Advanced Computer Science and Applications. 2026;17(10). doi:10.14569/ijacsa.2026.0171092
BibTeX
@article{Adnani2026,
author = {Adnani, El Mehdi and Erritali, Mohammed},
title = {Federated Credit Card Fraud Detection Under Simulated Operating Conditions: Encrypted Aggregation, Partial Participation and Byzantine Resilience},
journal = {International Journal of Advanced Computer Science and Applications},
year = {2026},
volume = {17},
number = {10},
doi = {10.14569/ijacsa.2026.0171092}
}
EUVIMED