Published Paper Details:

detecting tax fraud is the highest priority of almost all of the tax administrations aiming at the maximization of revenues and most importantly, the high level of compliance. These methods, including data mining, machine learning and the other methods, such as traditional random auditing, have already been applied to a large proportion of studies to address tax fraud. Work of this study is to apply Artificial Neural Networks in order to detect tax fraud factors in income tax data. The findings indicate that the Artificial Neural Networks show strong behaviors in the tax fraud detection with an accuracy at 92%, a precision at 85%, a recall score at 99% and an AUCROC value at 95%. All businesses, cross-border or domestic, the size of the business, small businesses or corporate businesses, are to be found among the factors considered by the model to be of greater saliency for income tax fraud detection. In this work, the paper is in agreement with the previous closely related to the previous tax fraud feature that covered all tax types jointly using various machine learning models. To our knowledge, this paper is the first to apply Artificial Neural Networks in the detection of income tax fraud in Rwanda, by different parameters such as layers, batch size, and epochs tuning, to select the most appropriate parameters which perform better than other parameters, in terms of accuracy. In this work, for this subject, it is found that with a simple model, no hidden layer, softsign activation function, performs more excellent. The findings of this work will assist auditors to get a grasp of the factors contributing to income tax fraud in order to decrease audit work generated, to decrease audit costs, and to recover the money lost due to an income tax fraud.