Application of unsupervised machine learning techniques for autonomous financial fraud detection in decentralized blockchain environments: a systematic review

Abstract

The rapid growth of decentralized blockchain-based financial systems has introduced new challenges in fraud detection, particularly due to anonymity, scalability, and the dynamic nature of transactions. Traditional supervised learning approaches often prove insufficient in these environments due to their reliance on labeled data and fixed patterns. This systematic literature review (SLR) investigates the application of unsupervised machine learning techniques for autonomous fraud detection in decentralized blockchain environments, analyzing their effectiveness, adaptability, and limitations compared to supervised or semi-supervised methods. Using a rigorous methodology based on the PICOC framework and PRISMA guidelines, we reviewed 20 studies published between 2021 and 2025. The results reveal that clustering and anomaly detection algorithms (such as autoencoders and graph-based methods) achieve superior performance (85–99% accuracy) in identifying common frauds like phishing and Ponzi schemes, leveraging blockchain's transparency and immutability. Key metrics such as recall (90–95%) and F1-score (88–93%) prove critical for evaluating models on imbalanced datasets. However, challenges persist in scalability, privacy, and cross-protocol adaptability. This study contributes a taxonomy of unsupervised techniques applied to blockchain fraud detection and proposes future research directions, such as hybrid models and standardized evaluation frameworks for decentralized ecosystems.