As technology advances and attack techniques show continuous progress, intrusion detection systems also need to develop. Hence, machine learning techniques started to be adopted more and more in the research on intrusion detection systems. In this study, application of Recurrent and Feed-forward Neural Network models is examined for detection and classification of attacks on computer networks. Using the dataset CICIDS2017 that has 14 up-to-date attack types, numerous tests are run with Multilayer Perceptron (MLP), Long-short Term Memory (LSTM), and Gated Recurrent Unit (GRU) neural network models and their success rates for the classification of malicious packets are examined. Additionally, based on the obtained results, a second set of tests are performed to observe the effect of the features in the dataset on success rates and performance. As a result of ablation study, performed by reducing the number of attributes used by one for each feature, the accuracy rates of different network models are obtained and it has been shown that using the most effective features, the duration of the tests can be reduced without affecting the detection rates significantly.