Predicting anomalies in computer networks using autoencoder-based representation learning

Abstract

Recent  improvements  in  the  internet  of  things  (IoT),  cloud  services,  and network  data  variety  have  increased  the  demand  for  complex  anomaly detection algorithms in  network  intrusion detection systems (IDSs)  capable of  dealing  with  sophisticated  network  threats.  Academics  are  interested  in deep  and  machine  learning  (ML)  breakthroughs  because  they  have  the potential   to   address   complex   challenges   such   as   zero-day   attacks.   In comparison  to  firewalls,  IDS  are  the  initial  line  of  network  security.  This study    suggests    merging    supervised    and    unsupervised    learning    in identification  systems  IDS.  Support  vector  machine  (SVM)  is  an  anomaly-based     classification     classifier.     Deep     autoencoder     (DAE)     lowers dimensionality.  DAE  are  compared  to  principal  component  analysis  (PCA) in  this  study,  and  hyper-parameters  for  F-1  micro  score  and  balanced accuracy  are  specified.  We  have  an  uneven  set  of  data  classes.  precision-recall  curves,  average  precision  (AP)  score,  train-test  times,  t-SNE,  grid search,   and   L1/L2   regularization   methods   are   used.   KDDTrain+   and KDDTest+  datasets  will  be  used  in  our  model.  For classification  and performance,   the   DAE+SVM   neural   network   technique   is   successful. Autoencoders outperformed linear PCA in terms of capturing valuable input attributes   using   t-SNE   to   embed   high   dimensional   inputs   on   a   two-dimensional  plane.  Our  neural  system  outperforms  solo  SVM  and  PCA encoded SVM in multi-class scenarios.