A GAN-Driven Approach for Anti-Malware System Design and Performance Analysis

Abstract

Malware detection continues to be a major challenge as attackers increasingly outperform traditional signature based defenses with novel evasion strategies. Classic techniques—such as pattern matching and anomaly detection—often fail against emerging or heavily obfuscated malware. To counter these threats, researchers are turning to Generative Adversarial Networks (GANs) as a proactive defense mechanism. In this setup, a generator network crafts synthetic, adversarial malware variants, while a discriminator network learns to differentiate between these and legitimate software. Training with these dynamically generated samples exposes the classifier to diverse, hard-to-detect threats, significantly improving its robustness. As demonstrated in approaches like Mal-LSGAN, such GAN augmented systems can maintain over 95% classification accuracy while successfully evading a wide range of baseline detectors Applying this concept to malware detection on Windows binaries using datasets from sources like VirusShare, the GAN-enhanced classifier achieved a detection accuracy of 94.6%, outperforming conventional methods. Evaluations focused on detection accuracy, false positive rate, and computational efficiency, showing that integrating GAN-generated adversarial samples boosts adaptability against constantly evolving threats. Moreover, this deep-learning-driven defense underscores the practical benefits of adversarial training in cybersecurity. By simulating future malware variants during training, the system stays ahead of attackers and strengthens current detection pipelines.