Exploring the effect of training-time randomness on the performance of deep neural networks for intrusion detection