A Machine Learning Approach of Predicting Ransomware Addresses in BlockChain Networks

Ransomware attacks, particularly those employing cryptocurrencies such as Bitcoin, pose a significant risk to worldwide cybersecurity. The decentralized structure of these networks, coupled with the anonymity they afford, complicates the tracking and cessation of criminal operations. This paper presents an extensive analysis focused on improving ransomware detection utilizing advanced machine learning and deep learning approaches on multiclass ransomware classification. Our investigation employed three principal models: Bidirectional Long Short-Term Memory (BiLSTM), Multi-Layer Perceptron (MLP), and Support Vector Machine (SVM). To address inherent class imbalances and enhance model efficacy, we incorporated advanced techniques, including Topological Data Analysis (TDA), Focal Loss, and Synthetic Data Generation. Topological Data Analysis (TDA) revealed complex data patterns that conventional approaches may overlook, while Focal Loss was integrated to emphasize difficult instances and address class imbalance. Moreover, Generative Adversarial Networks (GANs) produced synthetic data, improving dataset equilibrium and fortifying model training. In multiclass classification, the BiLSTM model emerged as the superior performer, attaining an accuracy of 85.54% and exceeding both MLP and SVM. This underscores the versatility and efficacy of BiLSTM across multiple prediction categories. Our research highlights the essential integration of TDA, GAN, and focal loss to improve feature representation and model resilience in ransomware detection.