A Model for Estimating Resiliency of AI-Based Classifiers Defending Against Cyber Attacks

Abstract

Abstract AI techniques for cybersecurity are advancing, but AI-based classifiers are suspectable of adversarial attacks. It is challenging to quantify the efforts required of an adversary to manipulate a system and quantify this resilience such that different systems can be compared using standard metrics. The study intends to quantify the actions required when an attacker abuses an AI-based system and propose a model to assess the attacker’s cybersecurity resilience. The study proposes an Egyptian Vulture Optimized Adaptive Elman Recurrent Neural Networks (EVO-AERNN) model to assess cybersecurity resilience and compare it with machine learning and deep learning-based classifiers. It illustrates the potential of using adversary-aware feature sampling to build more robust classifiers and use an optimized algorithm to maintain inherent resilience. The proposed model is achieved with an accuracy of 0.995, an F1 score of 0.9932, a precision of 0.9921, a recall (before an attack) of 0.987, a recall (after an attack) of 0.632, and a severity score of 0.363. The proposed model is further validated with a secondary dataset. This study paves the way for a more comprehensive knowledge of adversarial attack scenarios on network systems and offers valuable insights, inspiring further research on advancing cybersecurity studies.

Keywords

• Adversarial learning
• Resilience
• AI classifiers
• Cyber security