Syn-SWIFT: Synthetic SWIFT Transaction Dataset for Federated Fraud Detection

The inaccessibility of financial data constitutes a fundamental barrier to the development of robust AI models for frauddetection. Strict privacy regulations isolate repositories of data within an organization that cannot be accessed or shared by other departments, usually termed data silos, preventing institutions from pooling data for centralized training. This paper introduces Syn-SWIFT, a methodological framework for generating high-fidelity, synthetic, and labeled SWIFT message data. The framework is not a black-box generator; it is an explainable, hybrid system built on rigorously defined ground-truth constraints derived from expert domain knowledge, from 12-character BIC logic to incoming/outgoing message simulation to ensure all generated messages are semantically and structurally valid. The framework’s efficacy is demonstrated by generating four distinct datasets (MT101, MT103, MT110, MT202), totaling 400,000 transactions, segregated into 20 bank-specific silos. A modular pipeline incorporating fraud-pattern mutation, metadata synthesis, and raw message reconstruction ensures semantic correctness while maintaining operational variability. Each dataset is injected with a comprehensive lexicon of 40 unique, forensically sound fraud types, enabling complex multi-class classification. Experimental validation demonstrates the system’s effectiveness in producing coherent metadata–message alignment, stable reconstruction performance, and consistent fraud-type distribution across message types. Key challenges, including semantic drift, parser consistency, and multi-bank message fidelity, were addressed through rule-driven header enforcement and structured field regeneration. This siloed dataset is explicitly designed to serve as a new, public benchmark to accelerate research in Federated Learning (FL) for privacy-preserving financial crime detection.