ADCC-Bench: A benchmark framework for anomaly detection in cryptocurrency transactions
Author Identifier (ORCID)
Abstract
Anomaly detection in cryptocurrency transactions is critical for blockchain forensics, yet fragmented evaluation protocols and information leakage from non-chronological data splits undermine cross-study reliability. We present ADCC-Bench, a unified benchmarking framework built on three methodological controls: modality-aware data splitting that selects partitioning rules based on each datasef's temporal structure, standardized preprocessing to eliminate configuration bias across model families, and multi-seed replication (N=20) validated with Welch's t-tests to distinguish genuine performance differences from stochastic noise. Using this protocol, we evaluate tree-based models (Random Forest, XGBoost, LightGBM), graph neural networks (GCN, GraphSAGE), and their hybrid ensembles across three datasets representing distinct anomaly typologies: transaction-flow (Elliptic++), execution-level (BLTE), and behavioral (Ethereum), with SHAP-based feature attribution to verify that model decisions rely on domain-relevant forensic indicators. Our experiments yield three key findings: (i) random splitting inflates Macro-FI by up to 11% and reverses model rankings, confirming temporal leakage severity; (ii) tree-based models consistently achieve the highest single-model Macro-FI, while GCN-augmented ensembles provide complementary gains on transaction-flow data; and (iii) ensemble accuracy gains are not statistically significant; their primary benefit is variance reduction in noisy settings. ADCC-Bench establishes a reproducible, structure-aware benchmark for principled model selection in blockchain forensics.
Keywords
anomaly detection, benchmark framework, blockchain forensics, cryptocurrency, explainable artificial intelligence, graph neural networks
Document Type
Conference Proceeding
Date of Publication
1-1-2026
Publication Title
2026 International Wireless Communications and Mobile Computing (IWCMC)
Publisher
IEEE
School
School of Science
ISBN
[9798331550011]
Copyright
subscription content
Content Type
Metadata only
First Page
1501
Last Page
1506
Recommended Citation
Khanh, H. Q., Khoa, N. A., Huynh, V., Ahmed, M., To, T. N., Pham, V. H., & Duy, P. T. (2026). ADCC-Bench: A benchmark framework for anomaly detection in cryptocurrency transactions. In 2026 International Wireless Communications and Mobile Computing (IWCMC) (pp. 1501-1506). IEEE. https://doi.org/10.1109/IWCMC69287.2026.11580002