Nowadays, Botnet detection plays a vital role in assuring of information and internet security. This research introduces a scheme for peer-to-peer Botnet detection using a deep neural network in collaboration with the features selection approach. A classification and regression tree, ReliefF algorithm, and principal component analysis are utilized to choose the most significant features, and a deep neural network model is built based on adaptive learning methods (ADAM). The approach used in the proposed system utilizes network traffic alone, and the packet payload does not influence it, thus, avoids inherent shortcomings, such as the failure to handle encrypted payloads, as well as, preventing unknown malware from being addressed with rule-based antivirus software to be combated. This study compares the proposed model with classical machine learning methods like Gaussian NB, Logistic regression, SVM and Random Forest. The experiment results note that the proposed deep neural network model achieve the best performance with a principal component analysis as a features selection method. From the experimental results, the proposed method reached the accuracy of 98.2% along with 0.015 % false positive rate and 0.64 Root mean square error.