Author Identifier
Date of Award
2026
Keywords
cryptocurrency anomaly detection, blockchain temporal features, block height analysis, feature selection, supervised learning, unsupervised anomaly detection, explainable artificial intelligence, LIME
Document Type
Thesis
Publisher
Edith Cowan University
Degree Name
Master of Computing and Security by Research
School
School of Science
First Supervisor
Syed Mohammed Shamsul Islam
Second Supervisor
Iqbal Sarker
Third Supervisor
Mohiuddin Ahmed
Abstract
Blockchain-based cryptocurrencies enable decentralised and transparent financial transactions; however, their pseudonymous design has also facilitated a wide range of illicit activities, including fraud, money laundering, ransomware payments, and sanctions evasion. Although blockchain transaction records are publicly available, identifying anomalous or illicit wallet behaviour remains challenging due to extreme class imbalance, evolving adversarial strategies, and the limited interpretability of many machine learning models used in prior studies.
This thesis proposes a systematic and interpretable framework for cryptocurrency wallet anomaly detection that leverages blockchain-native temporal features derived from block height. Two complementary experimental investigations are conducted using the Elliptic++ dataset. The first investigation integrates multiple feature selection techniques with supervised learning models to assess classification performance and feature relevance under severe class imbalance, while incorporating local explainability through LIME. The second investigation performs a large-scale comparative evaluation of diverse unsupervised anomaly detection algorithms operating without access to class labels during training, with labels used exclusively for contamination estimation, threshold calibration, and post-hoc evaluation.
Experimental results demonstrate that supervised learning models achieved strong classification performance, with Random Forest and ExtraTrees obtaining accuracy values of up to 0.96 and F1-scores of up to 0.95 under selected feature configurations. In the unsupervised setting, HBOS and ECOD_like emerged as the strongest-performing anomaly detection models across temporal and full feature spaces, respectively. The results further show that non-temporal features achieve stronger predictive performance due to their statistical discriminative power, whereas temporal blockchain features provide substantially greater semantic interpretability by capturing wallet lifespan, activity regularity, and chronological behavioural patterns. Quantitative interpretability analysis additionally reveals a clear trade-off between predictive accuracy and explanatory transparency, emphasising the importance of temporal features for forensic and compliance-oriented blockchain analysis.
Overall, this thesis demonstrates that temporally aware feature analysis, systematic feature selection, and explainable machine learning can be integrated as complementary objectives rather than treated as competing goals. The proposed framework advances trustworthy and auditable cryptocurrency anomaly detection by balancing detection performance with behaviourally meaningful explanations suitable for real-world blockchain forensics.
Access Note
Access to this thesis is embargoed until 1st July 2027
Recommended Citation
Islam, M. (2026). An approach for identifying anomalous cryptocurrency transactions using temporal blockchain features and machine learning with explainability analysis. Edith Cowan University. https://doi.org/10.25958/q1za-zr37