DETECTION OF BITCOIN FRAUD USING VARIATIONAL AUTO ENCODER AND LIGHT GRADIENT BOOSTING MACHINE

The proliferation of fraudulent behaviour in cryptocurrency ecosystems poses a significant threat to financial security infrastructure. This study develops an approach that uses Variational Autoencoders in combination with Light Gradient Boosting Machines for detecting fraudulent Bitcoin transactions. The Variational Autoencoder component encodes each transaction into a compact latent representation, which is concatenated with the original transaction features to enrich the representation available to the classifier. The Light Gradient Boosting Machine is responsible for classification. Experimental validation on the Elliptic dataset of bitcoin transactions shows the model achieves 99.2% accuracy and 96% F1 score for minority-class detection, outperforming other methods such as Random Forest (95.9% F1), ANN-based models (89.8% F1), and ensemble stacking (97% F1). These results far exceed those of previous approaches, demonstrating the model’s ability to enhance the security and integrity of cryptocurrency transactions.