An Explainable Hybrid CNN–Transformer Framework for Fraud Detection in Digital and Cryptocurrency Transactions
Year: 2026
Type: article
Abstract: Fraud detection in digital payment and cryptocurrency systems is complicated by heterogeneous transaction records, class imbalance and limited model interpretability. This study proposed an explainable hybrid convolutional neural network (CNN)–Transformer framework for identifying suspicious transactions using the IEEE-CIS Fraud Detection and Ethereum Address Behavior datasets. Tabular features were transformed into spatial representations using the Image Generator for Tabular Data algorithm, while principal component analysis and rolling temporal sequences were employed to represent transaction behaviour. A dual-branch CNN–Transformer model was developed with cost-sensitive learning and hyperparameter optimisation using the Sparrow Search Algorithm. Model predictions were interpreted using SHapley Additive exPlanations, while an IPFS- and blockchain-based architecture was proposed for audit logging. Under the reported chronological evaluation procedure, the model achieved a precision of 0.7320, recall of 0.4956, F1-score of 0.5910 and area under the precision–recall curve of 0.5984 at a decision threshold of 0.5350. These results indicate moderate fraud-discrimination performance, although approximately half of the labelled fraudulent cases were not detected at the selected threshold. Further evaluation against established baseline models, independent validation of cryptocurrency fraud labels and experimental assessment of the proposed blockchain subsystem are necessary to establish the framework's comparative effectiveness and operational suitability.
Source: FUDMA Journal of Sciences
Institution: Al-Hikmah University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7226595933
