Improving Cyber Defense Against Ransomware: A Generative Adversarial Networks-Based Adversarial Training Approach for Long Short-Term Memory Network Classifier

• DOI: 10.3390/electronics14040810
• Publication Date: 2025-02-19T10:34:26Z
• Authors: Ping Wang; Hsiao-Chung Lin; Jia-Hong Chen; Wen-Hui Lin; Hao-Cyuan Li
• Keywords: Generative adversarial network; Adversarial machine learning; Ransomware

The rapid proliferation of ransomware variants necessitates more effective detection mechanisms, as traditional signature-based methods are increasingly inadequate. These conventional rely on manual feature extraction and matching, which time-consuming limited to known threats. This study addresses the escalating challenge threats in cybersecurity by proposing a novel deep learning model, LSTM-EDadver, leverages Generative Adversarial Networks (GANs) Carlini Wagner (CW) attacks enhance malware capabilities. LSTM-EDadver innovatively generates adversarial examples (AEs) using sequential features derived from behaviors, thus training models improve their robustness accuracy. methodology combines Cuckoo sandbox analysis with conceptual lattice ontology capture wide range families variants. approach not only shortcomings existing but also simulates real-world conditions during validation phase subjecting CW attacks. experimental results demonstrate that achieves classification accuracy 96.59%. performance was achieved dataset 1328 samples (across 32 families) 519 normal instances, outperforming RNN, LSTM, GCU models, recorded accuracies 90.01%, 93.95%, 94.53%, respectively. proposed model shows significant improvements F1-score, ranging 2.49% 6.64% compared without training. advancement underscores effectiveness integrating GAN-generated attack command sequences into