As the popularity of cryptocurrency, particularly Ethereum, as a mode of transaction increases, so does the emergence of fraudulent transactions as a threat to trade security. To protect users as well as administrators from partaking in a fraudulent transaction, identification of such transactions is necessary. This study aims to tackle this rising problem by proposing the application of Kolmogorov–Arnold Networks (KANs) and a weighted ensemble-network classifier and training them to detect fraudulent transactions on Ethereum smart contracts. In this paper, we apply the Kolmogorov–Arnold Networks as well as propose a 2-way soft-voting ensemble to identify fraudulent transactions on the Ethereum transaction dataset. Additionally, we use the Particle Swarm Optimizer with incorporated mutation (PSOM) for feature extraction and compare it against the features the model would generally learn without the algorithm. The performances of these models are comprehensively evaluated and statistically compared against approaches and commonly used algorithms. The KAN and the ensemble-model outperform the other approaches, with a k-fold test accuracy of 99.53% and 98.48% respectively, showing an improvement of about 0.93% from prior approaches with an inference time of 28 ms. This study enhances the detection and prevention of fraudulent transactions within the blockchain landscape, ultimately helping make transactions more secure and contributing to cybersecurity.