Many frauds have been experienced in recent years as a result of the existence of thousands of importers and exporters in each country and the difficulty in supervising them. These frauds have affected the tax which is the backbone of national revenue. As part of a solution, separated customs organizations of different countries including Ethiopia have been established to control foreign trade and takes legal action on the people and vehicles involved in the act of smuggling while it facilitates the legitimate movement of goods and people across the border. But customs organizations couldn’t protect the society significantly from adverse effects of smuggling. The reports from developing countries and independent organizations also have clear image for the increasing of customs transaction fraud. This circumstance, on the other hand, needs a daily thorough examination of a large volume of declarations, which consumes a significant amount of human resources and time. In view of this, deep learning based customs transaction fraud detection models have been developed to detect and categorize the fraud. The researchers used multi-layer perceptron ANN approach, back propagation algorithm in line with combinations of parameters using grid search hyper parameter tuning method to improve the performance of fraud detection and categorization. Past Ethiopian customs data of more than 95038 instances with 22 attributes have been used for model building. The fraud detection model and fraud category predictor model have been built with classification accuracy of 96.64% and 94.68% respectively. Keywords: ANN, MLP, ASYCUDA, fraud detection, fraud category, activation function.