Adaptive deanonymization of blockchain transactions leveraging integrative neural architectures in decentralized networks

Abstract The pseudonymous nature of decentralized networks like Bitcoin presents tremendous obstacles for regulatory compliance, forensic auditing, and cybercrime investigation. Traditional heuristics and single-modality machine learning (ML) frameworks provide partial transparency but usually suffer from the modality isolation problem, which fails to consider spatial graph structure, temporal transaction velocity, and local metadata patterns simultaneously. To break these limitations, in this paper, we propose an adaptive behavioral deanonymization method for blockchain transactions at the entity level based on an integrative neural architecture (CNN → LSTM → GNN) that dynamically focuses on spatial, temporal, and structural signals based on the behavior of the entity. The integrative neural uses a Disjoint Set Union (DSU) algorithm to combine temporary address artifacts into 140,235 resolved actor clusters, lowering ledger noise by 67.2%. The proposed architecture uses 1D-Convolutional Neural Networks (CNN) for extracting local structural fingerprints, Long Short-Term Memory (LSTM) networks for capturing temporal sequence velocity, and Graph Neural Networks (GNN) over a dynamic k-Nearest Neighbor (k-NN) adjacency matrix for aggregating neighborhood context. The architecture uses cost-sensitive learning (weighted cross-entropy) to tackle the significant class imbalance (1:500) to maintain analytical authenticity and forensics standards without synthetic distortion. The proposed architecture is assessed on a sample of a dataset of 427,625 transactions, achieving an overall classification accuracy of approximately 97%, which is better than the baseline models, and generating a transparent, explainable audit trail using SHAP and LIME feature attribution.