Automating Cyberdeception Evaluation with Deep Learning

Abstract

A machine learning-based methodology is proposed and implemented for conducting evaluations of cyberdeceptive defenses with minimal human involvement. This avoids impediments associated with deceptive research on humans, maximizing the efficacy of automated evaluation before human subjects research must be undertaken. Leveraging recent advances in deep learning, the approach synthesizes realistic, interactive, and adaptive traffic for consumption by target web services. A case study applies the approach to evaluate an intrusion detection system equipped with application-layer embedded deceptive responses to attacks. Results demonstrate that synthesizing adaptive web traffic laced with evasive attacks powered by ensemble learning, online adaptive metric learning, and novel class detection to simulate skillful adversaries constitutes a challenging and aggressive test of cyberdeceptive defenses.