anomaly-based intrusion detection using generelized variational auto-encoder

With the advent and increasing usage of the internet in the past few years, securing the exchange information among users is critical due to the growth of threats. considering this, researchers are focusing on firewalls and network intrusion detection systems (nidss), to detect and preventing malicious data and online attacks. since the firewalls are very expensive approaches, which barely adopt themselves with the fluid nature of the networks, nids are better solutions for protecting systems and information. vaes are powerful models on generalization and distribution learning which motivate us to propose a novel vae with two distribution learning layers with a new term of loss instead of mean square error (mse) for anomaly-based intrusion detection systems. we call this method as generalized vae (gvae) in the future. the nsl-kdd is used for evaluations and testing this model. we used only benign data for the training phase, and also used probability density function (pdf) on likelihood distribution which learned by vae for anomaly detection instead of mean square error, which is a more reliable measure than reconstruction mse for anomaly detection. also, the comparison between the gvae, simple vae, and auto-encoder has been provided.