Abstract: Now a days, the voice transmission technologies on the networks based on the IP (VOIP) has become one of the most widely used technologies in telecommunications due to lower costs and more flexibility. These networks due to variation of supporting VOIP terminals are vulnerable in terms of security and against several attacks such as lack of service, worms, DoS and DHCP attacks and etc. After attacking prevention methods, such as encryption, one of the conventional methods for securing VOIP is using Intrusion Detection Solutions (IDS) which include misuse detection methods and methods of detecting abnormal behavior. One area of concern in the designing the intrusion detection systems is machine learning and data mining. On the other hand, intrusion detection systems work with large volumes of data that include additional features and this, slow down and thus reduce the efficiency of the training and testing process. For this reason, feature selection, as one of the key issues of intrusion detection systems, includes finding a subset of more efficient features to improve the accuracy of prediction. In this study a method is proposed in four phases, to identify the best available features for intrusion detection and use them to design a compound classifier in order to detect the attack packets to the network. The results of the proposed simulation method is indicative of a 99.73% accuracy in detecting attacks.