Detecção de fraudes em criptomoedas utilizando métodos de classificação de séries temporais baseados em redes neurais

Abstract

This work presents a method of detecting fraud in cryptocurrencies, originated from an Initial Coin Offering (ICO). For this, predictive models were used, based on neural networks, for the classification of time series, generated from transaction flow tables in the Ethereum network. The first ICO activity was carried out in 2013 and reached its peak in the first half of 2018, with a turnover between 7 and 12 billion dollars worldwide. However, it is estimated that 78% of ICO activities are fraudulent. Based on the behavior of fraudulent and non-fraudulent cryptocurrencies, as well as on the transaction tables of the collected cryptocurrencies, 5 normalized time series were developed, which were input into the following types of Artificial Neural Networks (ANN): Multi Layer Perceptron (MLP); Convolution Neural Network - Multi Layer Perceptron (CNN-MLP); and Long Short Term Memory - Multi Layer Perceptron (LSTM-MLP) designed for classification. At the end of the research, a Recall value of up to 91% was obtained in some cases.