METHOD OF CLASSIFICATION OF TRANSACTIONS AND SEPARATE TRANSACTION PATTERNS BY NEURAL NETWORK MEANS ACCORDING TO SECURITY LEVEL IN BLOCKCHAIN SYSTEMS

Abstract

The paper develops a hardware-software method for classifying transactions and separate transaction patterns by neural network means according to the level of security in blockchain systems, focused on managing admission and pre-consensus processing modes without modifying the basic consensus mechanisms. The method operates in a dedicated off-chain trusted risk assessment loop and forms a transaction profile as a consistent representation of structural, temporal, and contextual features suitable for subsequent cryptographic fixation. A transformer-type neural network module calculates the transaction risk assessment, and the stability of the forecast is assessed using the Monte Carlo Dropout method, which provides controlled consideration of uncertainty as a parameter of the protocol policy. Based on the risk assessment and uncertainty, engineering zoning of transactions into “White”, “Gray,” and “Black” zones is performed with the formation of a protocol directive for standard processing, enhanced verification, or quarantine mode with delayed inclusion. Separately, the analysis of transaction patterns as recurring topological-temporal motifs in the transfer graph with the aggregation of their characteristics in a time window and the formation of pattern-oriented control signals is supported. To ensure the provability and reproducibility of the results, logging of decisions in the form of signed records with cryptographic commitments to features and cryptographic linking of records into a chain is implemented. Experimental verification was carried out on the "Elliptic Data Set" (Bitcoin) using a labeled subsample of 46,564 transactions and an extended profile of 188 features. The test values ​​​​of ROC-AUC 0.9207 and AUPRC 0.7588 were obtained; the zoning policy ensures the escalation of 0.8596 risky transactions to increased control modes at a load of 0.3155, and the overhead microbenchmark confirms a 57.98% reduction in the average cost of pre-consensus processing relative to solid enhanced verification. The practical value lies in implementing risk-based management of verification resources and audit decisions in a blockchain node.