Spiking neural networks (SNN) coupled with neuromorphic hardware are becoming popular for their extremely low energy benefits and applicability in compute & power constrained edge devices such as wearables, automated cars etc. In many implementations, surrogate gradient functions are used to apply classical back-propagation based learning on SNN to overcome the problem related to discontinuous nature of spikes. There are some variants of surrogates that are commonly used but there is no detailed study to understand their individual pros and cons. In this work, we have done one study using three such popular surrogates using a standard feed-forward SNN by taking the use case of network intrusion detection and attack classification at edge devices. For an SNN model that performs at par with a similar ANN, it is observed that continuous surrogates rather than discrete surrogates are more suited for better learning capability of the SNN. Additionally, networks using loss calculated with respect to the membrane potential of output neurons significantly outperform those using loss based on the number of output spikes for back-propagation.