Abstract

Currently, digital security mechanisms like Anomaly Detection Systems usingAutoencoders (AE) show great potential for bypassing problems intrinsic to thedata, such as data imbalance. Because AE use a non-trivial and nonstandardizedseparation threshold to classify the extracted reconstruction error, thedefinition of this threshold directly impacts the performance of the detectionprocess. Thus, this work proposes the automatic definition of this thresholdusing some machine learning algorithms. For this, three algorithms wereevaluated: the K-Nearst Neighbors, the K-Means and the Support Vector Machine.

 

Quick Read (beta)