Assessing of blockchain transactions using GATv2 and Graph-SMOTE
Radiotekhnika · 03.11.2026 · (226) · 76-84
DOI: 10.30837/rt.2026.3.226.06 ↗Kurz gefasst
Fachartikel in der Zeitschrift Radiotekhnika, online erschienen am 03.11.2026. Verfasst von V.V. Prosolov und 1 weiteren Autor:innen.
Abstract (Originalsprache)
This article addresses the scientific and applied problem of assessing transaction threats in blockchain systems under conditions of pseudo-anonymity, high feature sparsity, a significant proportion of unlabelled nodes, and extreme class imbalance. The starting premise of the study is that it is advisable to analyse a blockchain ledger not as a collection of isolated records, but as a directed attributed graph in which nodes correspond to transactions or addresses, edges reflect the movement of assets, and the feature matrix combines local transaction parameters with the topological characteristics of their surroundings. It is argued that common machine learning methods and standard SMOTE do not provide adequate compensation for imbalance in graph data, as they generate synthetic objects without accounting for transactional connectivity and may disrupt the structural semantics of the UTXO model. The proposed approach stays on the synthetic oversampling of the minority class is performed not in the primary tabular space, but in the latent space of graph embeddings formed by a GATv2 neural network with a dynamic attention mechanism. This approach allows to form the synthetic anomaly nodes while accounting for the graph’s multidimensional structural connectivity and preserving the model’s ability to aggregate the context of neighbouring nodes. The methodological article basis includes the combination of three main components: on-chain data graph representation, a GATv2 encoder for constructing latent representations, and topologically consistent oversampling of the minority class with subsequent use of a weighted loss function. A comparative analysis of the MLP, GCN, GATv2, and GATv2 with Graph-SMOTE models is presented, demonstrating an increase in the F1-score to 0.8948 and the preservation of high detection completeness for illicit transactions. The practical significance of the results lies in the possibility of applying the proposed approach in AML/CFT, digital forensics, and SOC monitoring to reduce the proportion of false-negative decisions and improve the quality of risk scoring for blockchain transaction graphs.
Zitieren
APA
Prosolov, V., & Khalimov, H. (2026). Assessing of blockchain transactions using GATv2 and Graph-SMOTE. Radiotekhnika, 76-84. https://doi.org/10.30837/rt.2026.3.226.06
Vancouver
Prosolov V, Khalimov H. Assessing of blockchain transactions using GATv2 and Graph-SMOTE. Radiotekhnika. 2026:76-84. doi:10.30837/rt.2026.3.226.06
BibTeX
@article{Prosolov2026,
author = {Prosolov, V.V. and Khalimov, H.Z.},
title = {Assessing of blockchain transactions using GATv2 and Graph-SMOTE},
journal = {Radiotekhnika},
year = {2026},
number = {226},
pages = {76--84},
doi = {10.30837/rt.2026.3.226.06}
}
EUVIMED