The Effectiveness of Zero-Day Attacks Data Samples Generated via GANs on Deep Learning Classifiers.

Area of Science:

• Cybersecurity
• Machine Learning
• Data Science

Background:

• The increasing digitization of services heightens the need for robust cybersecurity measures.
• Detecting zero-day vulnerabilities in new software and hardware is a significant challenge for cybersecurity professionals.
• Zero-day exploits pose immediate threats to end-users due to rapid exploitation by malicious actors.

Purpose of the Study:

• To propose a holistic methodology for generating realistic zero-day attack data.
• To evaluate a Neural Network (NN) detector trained with and without synthetic data.
• To assess the effectiveness of Generative Adversarial Networks (GANs) in creating zero-day attack datasets.

Main Methods:

• Designed and employed Generative Adversarial Networks (GANs) to synthetically generate zero-day attack data in tabular format.
• Utilized the Zero-Day GAN (ZDGAN) for dataset generation.
• Trained and evaluated a Neural Network classifier using both original and ZDGAN-augmented datasets.

Main Results:

• The ZDGAN achieved equilibrium after approximately 5000 iterations, producing synthetic data nearly identical to original samples.
• The NN model trained with ZDGAN-generated data demonstrated superior performance compared to a model trained solely on original data.
• The enhanced model achieved high validation accuracy and minimal validation loss in detecting zero-day attacks.

Conclusions:

• Synthetic data generation using ZDGAN is effective for augmenting cybersecurity datasets.
• Training NN models with GAN-generated data significantly enhances zero-day attack detection capabilities.
• The proposed methodology offers a promising approach to bolster cybersecurity defenses against emerging threats.