Year: 2023
Type: article
Abstract: Adversarial evasions represent contemporary challenges for applications relying on Machine Learning (ML). The susceptibility of traditional ML inference systems introduces vulnerabilities that make botnet detectors susceptible to attacks through adversarial examples. Complex AI models and sophisticated attack techniques can be used to generate the evasions. One of the potential sources of evasion assaults is generative AI models. The lack of data, which causes ML classifiers to train with a bias toward samples from the majority of classes, is a serious concern as well. This paper proposed a novel “Deep Reinforcement Learning based Evasion Generative Adversarial Network” (DRLEVGAN) to protect evasion attacks and retain semantics of the attack sample. The proposed model also tackles the issues of data imbalance, evasion awareness, and maintaining functionality in the context of synthetic botnet traffic generation. This model does not need adversarial training for the machine learning classifiers since it can act as an adversarial-aware botnet detection model. DRLEVGAN demonstrates superior performance when compared to similar models such as “Auxiliary Classifier GAN (ACGAN) and Evasion Generative Adversarial Network (EVGAN)”.
Institutions: COER University, Koneru Lakshmaiah Education Foundation
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7217247909