HybridFL: A Federated Learning Approach for Financial Crime Detection
Authors: Afsana Khan, Marijn ten Thij, Guangzhi Tang, Anna Wilbik
Organizations: Department of Advanced Computing Sciences, Maastricht University, Netherlands · Department of Cognitive Science and Artificial Intelligence, Tilburg University, Netherlands
Federated learning (FL) is a privacy-preserving machine learning paradigm that enables multiple parties to collaboratively train models on privately owned data without sharing raw information. While standard FL typically addresses either horizontal or vertical data partitions, many real-world scenarios exhibit a complex hybrid distribution. This paper proposes Hybrid Federated Learning (HybridFL) to address data split both horizontally across disjoint users and vertically across complementary feature sets. We evaluate HybridFL in a financial crime detection context, where a transaction party holds transaction-level attributes and multiple banks maintain private account-level features. By integrating horizontal aggregation and vertical feature fusion, the proposed architecture enables joint learning while strictly preserving data locality. Experiments on AMLSim and SWIFT datasets demonstrate that HybridFL significantly outperforms the transaction-only local model and achieves performance comparable to a centralized benchmark.
Figures & tables
Fig. 1: Illustration of the hybrid data partition. The transaction party holds transaction-level records, while Banks A and B each maintain private account-level data for their respective customers. Transactions link the sender and receiver accounts, creating vertical connections across institutions and horizontal separation within each bank.
Fig. 2: HybridFL training workflow: The transaction party acts as the active party, fusing embeddings received from the sender and receiver banks (passive parties) with its own transaction embedding to produce a prediction and compute the loss and gradients with respect to embeddings. Computed gradients are sent back to the corresponding banks for updating local sender and receiver models. Bank local models are periodically synchronized via Federated Averaging
∇θT=∑i∈B∇eiT∂θT∂eiT
Algorithm 1 Hybrid Federated Learning (HybridFL)
Dataset
Banks
Transactions
Positive:Negative Samples
AMLSim
2
63330
70:30
SWIFT
10
810563
903:1
TABLE I: Datasets used for the experiments
Dataset
Transaction Party Features
Bank: General Account Features
Sender-Specific Features
Receiver-Specific Features
AMLSim
Transaction amount Timestamp of transaction Transaction type Fraud label Identifiers: transaction ID sender ID receiver ID
Account type Account status Country of residence Gender of account holder Count of prior suspicious reports Initial deposit amount Age of account Identifiers: account ID, bank ID
No. of outgoing transactions Total amount sent by account Average amount sent per transaction Identifier: account ID (sender role)
No. of incoming transactions Total amount received by account Average amount received per transaction Identifier: account ID (receiver role)
SWIFT
Settlement amount Transaction chain length Settlement currency Fraud label Identifiers: transaction ID sender ID receiver ID
Account risk flags New account indicator Suspicious account indicator Identifiers: account ID, institution ID
No. of transactions sent Total amount sent by account Average amount per sent transaction Variation in sent transaction amounts Velocity of outgoing transactions (past 24h) Identifier: account ID (sender role)
No. of transactions received Total amount received by account Average amount per received transaction Variation in received transaction amounts Velocity of incoming transactions (past 24h) Identifier: account ID (receiver role)
TABLE II: Feature allocation across parties for AMLSim and SWIFT datasets
Model
Precision
Recall
F1 Score
AUPRC
Centralised
0.88
0.77
0.82
0.91
HybridFL
0.81
0.64
0.72
0.80
Transaction Party Only
0.39
0.50
0.44
0.53
TABLE III: Performance of the HybridFL algorithm on the AMLSim dataset. AUPRC reflects performance across all thresholds, while precision, recall, and F1 score are reported at a fixed threshold of 0.5.
Model
Precision
Recall
F1 Score
AUPRC
Centralised
1.00
0.82
0.90
0.84
HybridFL
0.92
0.75
0.83
0.78
Transaction Party Only
1.00
0.19
0.31
0.65
TABLE IV: Performance of the HybridFL algorithm on the SWIFT dataset. AUPRC reflects performance across all thresholds, while precision, recall, and F1 score are reported at a fixed threshold of 0.5.
Department of Computer Science, University of Bucharest · Luxembourg Institute of Science and Technology (LIST), Luxembourg · Research Institute of the University of Bucharest (ICUB), Bucharest, Romania