Money laundering is a serious worldwide criminal activity in financial institutions and a process in which illegitimate funds and assets are converted into legitimate funds and assets. These illegitimate funds can be found from serious crimes, such as drug trafficking, smuggling, political corruption, terrorist activity, financial fraud and so on. Suspicious money laundering transaction detection is very difficult when the dataset is imbalanced and the aim of this thesis work is, balancing the imbalance data using various techniques and increasing the performance accuracy of suspicious money laundering transactions detection using deep learning. The proposed algorithm evaluated by using an Elliptical dataset that has been developed for money laundering detection in 2019 due to the secrecy of Financial Intelegence Unit (FIU) data. This data set located from Kaggle.com and has 203,769 rows with 166 columns, from this amount of data two percent (4,545) are labeled class-1 (illicit). Twenty-one percent (42,019) are labelled class-2 (licit). The remaining transactions are not labeled whether licit or illicit but have other features. The dataset is highly imbalanced and to curve this problem, various data resampling techniques such as Random under-sampling, Random oversampling and Hybrid techniques have been done. In addition to this, we used Dropout and L2 regularization techniques during model training to constraint the model to fit our data accurately and not overfit. For experimental purpose we used python with keras library because we are familiar to this tool and it is currentstate- of-the-art. The proposed technique Convolutional Neural Network (CNN) in combination with different resampling techniques have been applied to detect whether the transactions are illicit or licit and the performance of the model is evaluated based accuracy result. The experimental result showed that the hybrid resampling technique so called Synthetic Minority Oversampling with Edited Nearest Neighbor (SMOTEENN) with L2 kernel regularize has a better result than the rest of resampling techniques and the accuracy improved from 96.1 to 98.93.