Generative Adversarial Network-Based Multi-Teacher Distilled Purification against Adversarial Malware Attacks

초록

The expansion of the attack surface due to the proliferation of cloud environments has driven a rapid increase in advanced cyber threats. Recently, visualization-based deep learning malware detection systems that convert input into grayscale images have been researched to counter these threats. However, deep learning models are vulnerable to adversarial attacks that mislead malware classifiers into classifying the attacks as benign due to their sensitivity to minor perturbations. Such vulnerability poses a critical security threat in cloud environments where accurate malware detection is vital for protecting sensitive data. Existing purification methods suffer from over-purification, which damages input data features, causing information loss. To address this problem, this study proposes a generative adversarial network-based multi-teacher distilled purification (GAN-MDPuri) scheme that distills knowledge from teacher models. The GAN-MDPuri scheme employs the competitive learning structure of GANs to achieve balanced knowledge distillation from a purification teacher network that removes perturbations and a reconstruction teacher network that restores normal sample input. Unlike existing knowledge distillation approaches that require complex weight adjustment mechanisms, the proposed scheme dynamically balances distinct knowledge during GAN training. The student model trained using this approach removes perturbations while preserving critical input features. The validation of the GAN-MDPuri scheme demonstrates that it achieved a high average adversarial malware detection accuracy of 0.9588 on large-scale datasets. Furthermore, the proposed model detected adversarial malware, reducing the adversarial attack success rate by an average of 0.3297 compared with existing purification schemes. This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/3.0/) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.

키워드