One-Class Classifier for Malicious Node Detection in Blockchain-Based IoT Networks

Blockchain is being explored as a substitute for central servers in IoT networks to enhance security by mitigating the limitations of the traditional architecture. But blockchain-based IoT networks are still prone to malicious nodes, which are attacked nodes in the network that transmit inaccurate data. The field of malicious node detection in blockchain-based IoT networks remains an area that requires further exploration. Hence, in this paper, we propose a system that uses One-Class Classification (OCC) algorithms for malicious node detection in blockchain-based IoT networks. Our system includes a Sensor Retention Policy (SRP) that decides whether a sensor should be removed from the network through trust points mechanics. 3 different OCC algorithms were considered for the system: (1) One-Class Support Vector Machine (OCSVM), (2) OCSVM using Standard Gradient Descent (SGD-OCSVM), and (3) Local Outlier Factor (LOF). Different variations of the system were created by varying the decision threshold parameter \(\varphi \) of each algorithm (the values considered are 0.0, \(-0.15\) , and \(-0.30\) ) and by also considering the system variation with no SRP. We devised a set of test runs to determine the performance of each system variation based on the following metrics: (1) Modified F-score, (2) Average Detection Time, (3) Average Processing Overhead, and (4) Memory Consumption. The test results show that using an OCSVM algorithm with \(\varphi = -0.30\) provides the highest recorded modified F-score of 0.98. Compared to the other OCC algorithms, the OCSVM algorithm is the most efficient in terms of the modified F-score and average detection time trade-off. The OCSVM algorithm provides the lowest average processing overhead (1.6 ns) and the lowest memory consumption (27.46 KB) compared to other OCC algorithms.