Ethiopian Ministry of Customs is the backbone of the country’s economy mandated to contribute to socio-economic development by facilitating trade and business across the border (Import and export) and on the other hand protecting national economies and societies against the threats posed by organized criminal syndicates and so on. The major objective of Customs authority i.e. Collection of Duties and taxes on international trade, protecting and combat smuggling of goods due to a poor implementation of customs risk management and customs clearance. One of the main issues during problem identification and understanding is the untracked and uncontrolled offense behavior of the declarant. Different types of fraudulent activities together with different result in iterative and frustrated declaration checkups inspections will result in assigning a significant number of inspectors and assessors. With an astronomical amount of data tends to be increase continually, together with frequent behavioral change and techniques to perform customs offense, poor data interpretation, by far not efficient enough to mitigate and solve the specified problem raised during the customs management process in general. This research proposes semi supervised deep learning technique, approach clearly classify, and ranking the illegal customs trade flows, which is the major contributor and decisive work to the overall customs revenue and duty when caught. A combination of a tree-based model for intelligible and transaction-level together with dual attention mechanisms is the robustness of our new model. Our model is able to identify evidences effectively that significantly determine whether a transaction is illicit. Moreover, the cross features with higher attention weights used to interpret the prediction results. Our model effectively predicts a precision of 92.66 % on illegally imported commodity transaction and a recall of 89.74% on revenue after inspecting after 10% of the imported inflow trade cargo. The model is vigorous against noise in the input data; clearly indicate any bias against Harmonized code as well as the country of the origin.