Anomaly Detection in Blockchain Transactions Using Isolation Forest and Autoencoder Deep Learning Models

Blockchain technology enables decentralized and transparent digital transactions, yet its open architecture also increases vulnerability to fraudulent and irregular activities. This study evaluates the effectiveness of the Isolation Forest method for detecting anomalous patterns within blockchain transaction data. A simulated dataset consisting of 10,130 transactions was constructed, including 62 injected anomalies that represent realistic irregular behaviours such as unusually large transaction values, extreme gas price spikes, and rapid consecutive transfers by a single sender. After applying feature engineering to capture temporal frequency, transaction dynamics, sender and receiver behaviour, and gas-related attributes, the Isolation Forest model was trained and evaluated using the embedded anomaly labels. The model achieved a precision of 0.4516, a recall of 0.4516, and an F1 score of 0.4516, indicating moderate detection capability. Analysis of the confusion matrix and anomaly score distribution further revealed overlapping characteristics between rare but legitimate transactions and true anomalies, which contributed to misclassification. Overall, the findings suggest that Isolation Forest can serve as an early anomaly filtering mechanism, although additional contextual information or hybrid detection strategies are needed to enhance performance in real blockchain environments.