Federated aggregation strategies for credit card fraud detection under differential privacy

Öz

Context—Card fraud cost the global financial system over $28 billion in 2019, and losses have risen every year since. Banks hold the transaction data needed for collaborative fraud detection, but privacy regulations such as GDPR and KVKK prevent cross-institutional data sharing. Federated Learning keeps raw data local: each institution trains on its own records and sends only model parameters to a shared coordinator. Differentially-Private Stochastic Gradient Descent (DP-SGD) counters gradient inversion attacks by clipping per-sample gradients and adding calibrated noise before parameters leave the client, yielding a record-level (ε, δ) guarantee whose strength depends on the accumulated privacy budget ε. How the aggregation strategy behaves under this noise has not yet been studied.

Objective—We compare five aggregation strategies—FedAvg, FedProx, cosine-similarity aggregation, FedAvg-DWA, and FedAdam—across two real-world benchmarks, two heterogeneity levels, four DP-SGD noise multipliers, and 10 random seeds per configuration. We ask whether the relative ranking of strategies survives DP noise, and what DP actually costs in deployed model behavior.

Method—510 federated training runs cover five aggregation configurations, two datasets (Kaggle ULB: n = 284807, 0.17% positive rate; IEEE-CIS: n = 590540, 3.50%), two Dirichlet non-IID levels (α ∈ {0.5, 0.1}), and σ ∈ {0.5, 1.0, 1.5, 2.0}. The base learner is a three-layer MLP with LayerNorm. Privacy accounting uses the Opacus RDP accountant at δ = 10⁻⁵. Performance is reported at both the fixed 0.5 threshold and a validation-tuned F₁-maximizing threshold. All pairwise comparisons carry Bonferroni-, Holm-, and Benjamini–Hochberg-corrected p-values, and all 200 realized client-shard partitions are characterized directly.

Results—On Kaggle ULB under IID partitioning, federated F₁@best-t (0.801–0.804) matches the centralized baseline (0.799 ± 0.036). Under mild non-IID without DP (α = 0.5), FedAvg-DWA narrowly leads (< 0.015 on F₁@best-t), though neither test survives correction. In the ULB privacy sweep, 7 of 40 Wilcoxon tests show nominal significance but none survive family-wise correction; performance consistently orders with FedAvg-DWA leading and FedAdam trailing. On IEEE-CIS, this ordering amplifies, and the extreme pair remains Holm-significant under DP (adjusted p = 0.020). While AUC holds near 0.95, calibration drift collapses precision at the fixed 0.5 threshold; a validation-tuned threshold recovers F₁ loss without privacy cost. Both calibration effects replicate on IEEE-CIS (ɛ = 8.10 ± 5.30 at σ = 1.0).

Conclusion—Aggregation-rule differences under DP-SGD are small and dataset-dependent in magnitude, consistent in direction, and on the ULB sweep none survives correction; outcome differences arise mainly from post-training threshold tuning. We also uncover a previously unreported interaction: Opacus's DP step quietly overrides the standard loss-augmented FedProx setup, making FedProx numerically equivalent to FedAvg, verified by bit-identical per-round global weight trajectories across seeds.

Anahtar Kelimeler

• Aggregation strategies
• Credit card fraud detection
• Differential privacy
• Federated learning
• Threshold calibration

Diferansiyel gizlilik altında kredi kartı dolandırıcılığı tespitine yönelik federe birleştirme stratejileri

Öz

Arka Plan—2019 yılında kredi kartı dolandırıcılığı küresel finansal sisteme 28 milyar dolardan fazla zarar vermiş; kayıplar o tarihten bu yana her yıl artmıştır. Bankalar, daha etkili dolandırıcılık tespit sistemleri geliştirmek için gerekli işlem verilerine sahiptir; ancak GDPR, KVKK ve benzer düzenlemeler kurumlar arası veri paylaşımını engellemektedir. Federe Öğrenme (FL), ham verileri yerel tutarak bu soruna pratik bir çözüm sunar. Federe öğrenme sürecinde her kurum modeli kendi kayıtlarıyla eğitir ve yalnızca model parametrelerini ortak bir koordinatöre iletir. Diferansiyel Gizliliğe Dayalı Rassal Gradyan İnişi (DP-SGD), gradyanlara iletimden önce kalibreli gürültü ekleyerek tersine mühendislik saldırılarına karşı biçimsel bir gizlilik güvencesi sağlar. Bununla birlikte, literatürde bu gürültü ortamında birleştirme stratejisi seçiminin etkisi henüz sistematik biçimde incelenmemiştir.

Amaç—Bu çalışmada beş federe öğrenme birleştirme stratejisi—FedAvg, FedProx, kosinüs-benzerlik birleştirme, FedAvg-DWA ve FedAdam—iki gerçek dünya kıyaslama veri kümesi, iki heterojenlik düzeyi, dört DP-SGD gürültü çarpanı ve yapılandırma başına 10 rastgele başlangıç noktası üzerinden karşılaştırılmaktadır. Temel hedef, stratejilerin göreli başarı sıralamasının DP gürültüsü devreye girdiğinde korunup korunmadığını belirlemek ve DP'nin konuşlandırılan model davranışına gerçek maliyetini ortaya koymaktır.

Yöntem—510 federe eğitim koşumu; beş birleştirme yapılandırmasını, iki veri kümesini (Kaggle ULB: n = 284807, %0.17 pozitif oran; IEEE-CIS: n = 590540, %3.50), iki Dirichlet IID-dışı düzeyini (α ∈ {0.5, 0.1}) ve σ ∈ {0.5, 1.0, 1.5, 2.0} gürültü çarpanlarını kapsamaktadır. Temel öğrenici LayerNorm içeren üç katmanlı bir MLP'dir. Gizlilik hesabı, δ = 10⁻⁵ ile Opacus RDP sayacı üzerinden yürütülmüştür. Performans, hem sabit 0.5 eşiğinde hem de doğrulama kümesi üzerinde F₁'i en büyükleyen ayarlı eşikte raporlanmıştır.

Bulgular—Hem bağımsız hem de özdeş dağılıma sahip (IID) bölümleme altında federasyonun performans maliyeti yoktur: federe F₁@en-iyi-t (0.801–0.804), merkezi taban çizgisiyle (0.799 ± 0.036) örtüşmektedir. DP olmaksızın hafif IID-dışı koşulda (α = 0,5), FedAvg-DWA dar bir farkla öne çıkmaktadır (F₁@en-iyi-t < 0.015); bu durum p = 0.049 ve p = 0.037 değerleriyle iki eşleştirilmiş Wilcoxon testiyle doğrulanmıştır. DP-SGD etkinleştirildiğinde ise gürültü düzeyinden bağımsız olarak 40 ikili Wilcoxon testinin tamamı p > 0.05 döndürmektedir. AUC süreç boyunca 0.95 düzeyinde seyrederken kalibrasyon sapması nedeniyle sabit 0.5 eşiğinde hassasiyet çökmektedir. Doğrulama kümesinde ayarlanan eşik, ek gizlilik maliyeti olmaksızın F₁ kayıplarının büyük bölümünü telafi etmektedir. Her iki etki IEEE-CIS üzerinde de doğrulanmıştır (σ = 1.0'da ɛ = 8.10 ± 5.30).

Sonuç—DP-SGD etkinleştirildiğinde birleştirme kuralı seçiminin nihai sonucu artık etkilemediği ortaya çıkmıştır; performans farklılıkları eğitim sonrası eşik ayarlamasından kaynaklanmaktadır. Bunun yanı sıra, daha önce bildirilmemiş bir etkileşim de tespit edilmiştir: Opacus'un DP adımı, kayıp tabanlı FedProx yapılandırmasını sessizce geçersiz kılarak FedProx'u sayısal olarak FedAvg ile eşdeğer hale getirmektedir.

Anahtar Kelimeler

• Birleştirme stratejileri
• Kredi kartı dolandırıcılığı tespiti
• Diferansiyel gizlilik
• Federe öğrenme
• Eşik kalibrasyonu

Kaynakça

1. E. Ileberi, Y. Sun, Z. Wang, “Performance evaluation of machine learning methods for credit card fraud detection using SMOTE and AdaBoost”, IEEE Access, 9, 165286–165294, 2021. https://doi.org/10.1109/ACCESS.2021.3134330.
2. A. Ali, S. A. Razak, S. H. Othman, T. A. E. Eisa, A. Al-Dhaqm, M. Nasser, T. Elhassan, H. Elshafie, A. Saif, “Financial fraud detection based on machine learning: A systematic literature review”, Applied Sciences, 12(19), 9637, 2022. https://doi.org/10.3390/app12199637.
3. European Parliament and Council of the European Union, “Regulation (EU) 2016/679 (General Data Protection Regulation)”, Official Journal of the European Union, L119, 1–88, 2016. https://gdpr-info.eu/.
4. H. B. McMahan, E. Moore, D. Ramage, S. Hampson, B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data”, Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS), Fort Lauderdale, FL, USA, 20-22 April 2017, 1273–1282. https://doi.org/10.48550/arXiv.1602.05629.
5. M. Abadi, A. Chu, I. Goodfellow, H. B. McMahan, I. Mironov, K. Talwar, L. Zhang, “Deep learning with differential privacy”, Proceedings of the 23rd ACM Conference on Computer and Communications Security (CCS), Vienna, Austria, 24-28 October 2016, 308–318. https://doi.org/10.1145/2976749.2978318.
6. C. Dwork, A. Roth, “The algorithmic foundations of differential privacy”, Foundations and Trends in Theoretical Computer Science, 9(3–4), 211–487, 2014. https://doi.org/10.1561/0400000042.
7. N. R. Shanbhog, K. S. Totad, A. R. Hanchinal, A. P. Bidargaddi, “Fraud detection in financial transactions using deep learning: A comparative study”, Proceedings of the 5th International Conference on Emerging Technologies (INCET), Belgaum, India, 24-26 May 2024. https://doi.org/10.1109/INCET61516.2024.10593486.
8. F. Z. El Hlouli, J. Riffi, M. A. Mahraz, A. El Yahyaouy, H. Tairi, “Credit card fraud detection based on multilayer perceptron and extreme learning machine architectures”, Proceedings of the International Conference on Intelligent Systems and Computer Vision (ISCV), Fez, Morocco, 09-11 June 2020, 1–5. https://doi.org/10.1109/1ISCV49265.2020.9204185.

9. S. K. Hashemi, S. L. Mirtaheri, S. Greco, “Fraud detection in banking data by machine learning techniques”, IEEE Access, 11, 3034–3043, 2023. https://doi.org/10.1109/ACCESS.2022.3232287.
10. K. Randhawa, C. K. Loo, M. Seera, C. P. Lim, A. K. Nandi, “Credit card fraud detection using AdaBoost and majority voting”, IEEE Access, 6, 14277–14284, 2018. https://doi.org/10.1109/ACCESS.2018.2806420.
11. D. Varmedja, M. Karanovic, S. Sladojevic, M. Arsenovic, A. Anderla, “Credit card fraud detection — Machine learning methods”, Proceedings of the 18th International Symposium INFOTEH-JAHORINA, East Sarajevo, Bosnia and Herzegovina, 20-22 March 2019, 1–5. https://doi.org/10.1109/INFOTEH.2019.8717766.
12. Y. Tang, Y. Liang, “Credit card fraud detection based on federated graph learning”, Expert Systems with Applications, 256, 124979, 2024. https://doi.org/10.1016/j.eswa.2024.124979.
13. T. K. Dang, T. Ha, “A comprehensive fraud detection for credit card transactions in federated averaging”, SN Computer Science, 5(5), 578, 2024. https://doi.org/10.1007/s42979-024-02898-y.
14. N. F. Aurna, M. D. Hossain, Y. Taenaka, Y. Kadobayashi, “Federated learning-based credit card fraud detection: Performance analysis with sampling methods and deep learning algorithms”, Proceedings of the IEEE Cyber Security and Resilience (CSR), Venice, Italy, 31 July-02 August 2023, 180–186. https://doi.org/10.1109/CSR57506.2023.10224978.
15. T. Baabdullah, A. Alzahrani, D. B. Rawat, C. Liu, “Efficiency of federated learning and blockchain in preserving privacy and enhancing the performance of credit card fraud detection systems”, Future Internet, 16(6), 196, 2024. https://doi.org/10.3390/fi16060196.
16. A. A. Süzen, B. Duman, “Blockchain-based secure credit card storage system for e-commerce”, Sakarya University Journal of Computer and Information Sciences, 4(2), 204–215, 2021. https://doi.org/10.35377/saucis.04.02.895764.
17. S. Alaklabi, “PrivChain-AI leveraging blockchain and federated learning for private financial reporting and access control”, Scientific Reports, 16(1), 2786, 2026. https://doi.org/10.1038/s41598-025-32606-6.
18. K. Bian, H. Zheng, “FedAvg-DWA: A novel algorithm for enhanced fraud detection in federated learning environment”, Proceedings of the 4th International Conference on Big Data, Artificial Intelligence and Internet of Things Engineering (ICBAIE), Hangzhou, China, 25-27 August 2023, 13–17. https://doi.org/10.1109/ICBAIE59714.2023.10281317.
19. S. Reddi, Z. Charles, M. Zaheer, Z. Garrett, K. Rush, J. Konečný, S. Kumar, H. B. McMahan, “Adaptive federated optimization”, arXiv, 2021, https://arxiv.org/abs/2003.00295.
20. T. Li, A. K. Sahu, M. Zaheer, M. Sanjabi, A. Talwalkar, V. Smith, “Federated optimization in heterogeneous networks”, arXiv, 2020. https://doi.org/10.48550/arXiv.1812.06127.
21. M. Nergiz, “Federe Öğrenmede Birleştirme Algoritmalarının Model Performansına Etkisi”, Dicle Üniversitesi Mühendislik Fakültesi Mühendislik Dergisi, 14(1), 65–73, 2023. https://doi.org/10.24012/dumf.1241947.
22. R. A. Mohammed, K. W. Wong, M. F. Shiratuddin, X. Wang, “Scalable machine learning techniques for highly imbalanced credit card fraud detection: A comparative study”, Proceedings of the 15th Pacific Rim International Conference on Artificial Intelligence (PRICAI) / 15th Pacific Rim Knowledge Acquisition Workshop (PKAW), Nanjing, China, 28-31 August 2018, 237–246. https://doi.org/10.1007/978-3-319-97310-4_27.
23. M. Alamri, M. Ykhlef, “Survey of credit card anomaly and fraud detection using sampling techniques”, Electronics, 11(23), 4003, 2022. https://doi.org/10.3390/electronics11234003.
24. N. V. Chawla, K. W. Bowyer, L. O. Hall, W. P. Kegelmeyer, “SMOTE: Synthetic minority over-sampling technique”, Journal of Artificial Intelligence Research, 16, 321–357, 2002. https://doi.org/10.1613/jair.953.
25. G. Lemaître, F. Nogueira, C. K. Aridas, “Imbalanced-learn: A Python toolbox to tackle the curse of imbalanced datasets in machine learning”, Journal of Machine Learning Research, 18, 17, 2017.
26. M. S. Farooq, S. F. Munir, M. F. Manzoor, M. Shaheen, “AI-driven adaptive federated learning with privacy preservation and imbalance adjustment for financial credit card fraud detection”, Applied Computational Intelligence and Soft Computing, 2025(1), 7116768, 2025. https://doi.org/10.1155/acis/7116768.
27. I. Mironov, “Rényi differential privacy”, Proceedings of the 30th IEEE Computer Security Foundations Symposium (CSF), Santa Barbara, CA, USA, 21-25 August 2017, 263–275. https://doi.org/10.1109/CSF.2017.11.
28. A. Yousefpour, et aI., “Opacus: User-friendly differential privacy library in PyTorch”, arXiv, 2021, https://arxiv.org/abs/2109.12298.
29. H. B. McMahan, D. Ramage, K. Talwar, L. Zhang, “Learning differentially private recurrent language models”, arXiv, 2018. https://doi.org/10.48550/arXiv.1710.06963.
30. K. Bonawitz, V. Ivanov, B. Kreuter, A. Marcedone, H. B. McMahan, S. Patel, D. Ramage, A. Segal, K. Seth, “Practical secure aggregation for privacy-preserving machine learning”, Proceedings of the ACM SIGSAC Conference on Computer and Communications Security (CCS), Dallas, TX, USA, 30 October - 03 November 2017, 1175–1191. https://doi.org/10.1145/3133956.3133982.
31. Machine Learning Group – ULB, “Credit card fraud detection”, Kaggle, 2013, https://www.kaggle.com/datasets/mlg-ulb/creditcardfraud, (15.04.2026).
32. IEEE Computational Intelligence Society and Vesta Corporation, “IEEE-CIS Fraud Detection”, Kaggle, 2019, https://www.kaggle.com/competitions/ieee-fraud-detection, (15.04.2026).
33. I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, Y. Bengio, “Generative adversarial networks”, Communications of the ACM, 63(11), 139–144, 2020. https://doi.org/10.1145/3422622.
34. A. Sharma, P. K. Singh, R. Chandra, “SMOTified-GAN for class imbalanced pattern classification problems”, IEEE Access, 10, 30655–30665, 2022. https://doi.org/10.1109/ACCESS.2022.3158977.
35. A. Paszke, et al., “PyTorch: An imperative style, high-performance deep learning library”, Proceedings of the 33rd Conference on Neural Information Processing Systems (NeurIPS), Vancouver, Canada, 08-14 December 2019.
36. D. J. Beutel, et al., “Flower: A friendly federated learning research framework”, arXiv, 2020, https://arxiv.org/abs/2007.14390.
37. T.-M. H. Hsu, H. Qi, M. Brown, “Measuring the effects of non-identical data distribution for federated visual classification”, arXiv, 2019, https://arxiv.org/abs/1909.06335.