The good, the bad, and the algorithm: The impact of generative AI on cybersecurity
Luigi Coppolino, Salvatore D’Antonio, Giovanni Mazzeo, Federica Uccello
Abstract
Generative Adversarial Networks (GANs) are emerging as a transformative technology in cybersecurity, presenting both opportunities and challenges in enhancing defensive and offensive strategies. This paper explores the impact that Generative Artificial Intelligence (AI) has on cybersecurity, focusing on its application in the field of network and web security. Current research reveals robust defensive approaches; however, there remains a significant gap in the application of Generative AI to develop advanced attack scenarios capable of bypassing existing defense mechanisms. Our work attempts to fill this gap and spreads awareness regarding a potential exposure of Neural Network (NN)-based Intrusion Detection Systems (IDSs) against AI-enhanced attacks. Unlike conventional approaches that focus on Input Perturbation, Data Poisoning, or Spoofing, we propose a novel offensive strategy called Attack Obfuscation. This strategy leverages Conditional GANs (CGANs) to conceal genuine attacks by injecting synthetic traffic designed to deceive NN-based IDS. The experimental investigation validates the proposed approach against three distinct datasets and different typologies of attacks, managing to successfully deceive the IDS.