Improving Fraud Detection with 1D-Convolutional Spiking Neural Networks Through Bayesian Optimization

The digitalization of the banking sector has enabled an increasing number of fraudulent activities in the past years. The development of new practical solutions for fraud detection is particularly challenging since the industry needs to respect a 5% threshold of false detection of fraud, avoiding monetary losses. The usage of traditional machine learning algorithms faces other challenges, such as classification discrimination and high energy consumption. Spiking Neural Networks, designed to mimic the brain’s natural processes, present a compelling alternative. Particularly on neuromorphic hardware, they exhibit remarkable energy efficiency, paving the way for the advancement of sustainable Artificial Intelligence. Our approach consists of applying 1D-Convolutional Spiking Neural Networks to improve fraud detection on the Bank Account Fraud dataset with the Bayesian optimization of hyperparameters. As a result, we show that these architectures can solve the problem of biased data while maintaining algorithm performance when compared against more traditional approaches such as Gradient Boosting models. Moreover, we lay out the Bayesian optimization process for the hyperparameters, thereby enhancing the performance of these algorithms.